Text book 2 computer networks_a systems approach_peterson solution manual

Computer Networks: A Systems Approach
Fifth Edition
Solutions Manual
Larry Peterson and Bruce Davie
2011
1

Dear Instructor:
This Instructors’ Manual contains solutions to most of the exercises in the fifth edition
of Peterson and Davie’s Computer Networks: A Systems Approach.
Exercises are sorted (roughly) by section, not difficulty. While some exercises are
more difficult than others, none are intended to be fiendishly tricky. A few exercises
(notably, though not exclusively, the ones that involve calculating simple probabilities)
require a modest amount of mathematical background; most do not. There is a sidebar
summarizing much of the applicable basic probability theory in Chapter 2.
An occasional exercise is awkwardly or ambiguously worded in the text. This manual
sometimes suggests better versions; also see the errata at the web site.
Where appropriate, relevant supplemental files for these solutions (e.g. programs) have
been placed on the textbook web site, http://mkp.com/computer-networks.
Useful other material can also be found there, such as errata, sample programming
assignments, PowerPoint lecture slides, and EPS figures.
If you have any questions about these support materials, please contact your Morgan
Kaufmann sales representative. If you would like to contribute your own teaching
materials to this site, please contact our Associate Editor David Bevans,
D.Bevans@elsevier.com.
We welcome bug reports and suggestions as to improvements for both the exercises
and the solutions; these may be sent to netbugsPD5e@elsevier.com.
Larry Peterson
Bruce Davie
March, 2011

Chapter 1 1
Solutions for Chapter 1
3. We will count the transfer as completed when the last data bit arrives at its desti-
nation. An alternative interpretation would be to count until the last ACK arrives
back at the sender, in which case the time would be half an RTT (25 ms) longer.
(a) 2 initial RTT’s (100ms) + 1000KB/1.5Mbps (transmit) + RTT/2 (propaga-
tion = 25ms)
≈ 0.125 + 8Mbit/1.5Mbps = 0.125 + 5.333 sec = 5.458 sec. If we pay
more careful attention to when a mega is 106
versus 220
, we get
8,192,000bits/1,500,000bps = 5.461sec, for a total delay of 5.586 sec.
(b) To the above we add the time for 999 RTTs (the number of RTTs between
when packet 1 arrives and packet 1000 arrives), for a total of 5.586 +
49.95 = 55.536.
(c) This is 49.5 RTTs, plus the initial 2, for 2.575 seconds.
(d) Right after the handshaking is done we send one packet. One RTT after the
handshaking we send two packets. At n RTTs past the initial handshaking
we have sent 1 + 2 + 4 + · · · + 2n
= 2n+1
− 1 packets. At n = 9 we have
thus been able to send all 1,000 packets; the last batch arrives 0.5 RTT later.
Total time is 2+9.5 RTTs, or .575 sec.
4. The answer is in the book.
5. Propagation delay is 4×103
m/(2×108
m/s) = 2×10−5
sec = 20 µs. 100 bytes/20 µs
is 5 bytes/µs, or 5 MBps, or 40 Mbps. For 512-byte packets, this rises to 204.8 Mbps.
7. Postal addresses are strongly hierarchical (with a geographical hierarchy, which
network addressing may or may not use). Addresses also provide embedded
“routing information”. Unlike typical network addresses, postal addresses are
long and of variable length and contain a certain amount of redundant informa-
tion. This last attribute makes them more tolerant of minor errors and inconsis-
tencies. Telephone numbers, at least those assigned to landlines, are more sim-
ilar to network addresses: they are (geographically) hierarchical, fixed-length,
administratively assigned, and in more-or-less one-to-one correspondence with
nodes.
8. One might want addresses to serve as locators, providing hints as to how data
should be routed. One approach for this is to make addresses hierarchical.
Another property might be administratively assigned, versus, say, the factory-
assigned addresses used by Ethernet. Other address attributes that might be
relevant are fixed-length v. variable-length, and absolute v. relative (like file
names).

Chapter 1 2
If you phone a toll-free number for a large retailer, any of dozens of phones may
answer. Arguably, then, all these phones have the same non-unique “address”. A
more traditional application for non-unique addresses might be for reaching any
of several equivalent servers (or routers). Non-unique addresses are also useful
when global reachability is not required, such as to address the computers within
a single corporation when those computers cannot be reached from outside the
corporation.
9. Video or audio teleconference transmissions among a reasonably large number
of widely spread sites would be an excellent candidate: unicast would require a
separate connection between each pair of sites, while broadcast would send far
too much traffic to sites not interested in receiving it. Delivery of video and audio
streams for a television channel only to those households currently interested in
watching that channel is another application.
Trying to reach any of several equivalent servers, each of which can provide the
answer to some query, would be another possible use, although the receiver of
many responses to the query would need to deal with the possibly large volume
of responses.
10. STDM and FDM both work best for channels with constant and uniform band-
width requirements. For both mechanisms bandwidth that goes unused by one
channel is simply wasted, not available to other channels. Computer communi-
cations are bursty and have long idle periods; such usage patterns would magnify
this waste.
FDM and STDM also require that channels be allocated (and, for FDM, be as-
signed bandwidth) well in advance. Again, the connection requirements for com-
puting tend to be too dynamic for this; at the very least, this would pretty much
preclude using one channel per connection.
FDM was preferred historically for TV/radio because it is very simple to build
receivers; it also supports different channel sizes. STDM was preferred for voice
because it makes somewhat more efficient use of the underlying bandwidth of
the medium, and because channels with different capacities was not originally
an issue.
11. 10 Gbps = 1010
bps, meaning each bit is 10−10
sec (0.1 ns) wide. The length in
the wire of such a bit is .1 ns × 2.3 × 108
m/sec = 0.023 m or 23mm
12. x KB is 8 × 1024 × x bits. y Mbps is y × 106
bps; the transmission time would
be 8 × 1024 × x/y × 106
sec = 8.192x/y ms.
13. (a) The minimum RTT is 2 × 385, 000, 000 m / 3×108
m/s = 2.57 seconds.
(b) The delay×bandwidth product is 2.57 s×1 Gbps = 2.57Gb = 321 MB.
(c) This represents the amount of data the sender can send before it would be
possible to receive a response.

Chapter 1 3
(d) We require at least one RTT from sending the request before the first bit
of the picture could begin arriving at the ground (TCP would take longer).
25 MB is 200Mb. Assuming bandwidth delay only, it would then take
200Mb/1000Mbps = 0.2 seconds to finish sending, for a total time of 0.2 +
2.57 = 2.77 sec until the last picture bit arrives on earth.
15. (a) Delay-sensitive; the messages exchanged are short.
(b) Bandwidth-sensitive, particularly for large files. (Technically this does pre-
sume that the underlying protocol uses a large message size or window size;
stop-and-wait transmission (as in Section 2.5 of the text) with a small mes-
sage size would be delay-sensitive.)
(c) Delay-sensitive; directories are typically of modest size.
(d) Delay-sensitive; a file’s attributes are typically much smaller than the file
itself.
16. (a) On a 100 Mbps network, each bit takes 1/108
= 10 ns to transmit. One
packet consists of 12000 bits, and so is delayed due to bandwidth (serial-
ization) by 120 µs along each link. The packet is also delayed 10 µs on
each of the two links due to propagation delay, for a total of 260 µs.
(b) With three switches and four links, the delay is
4 × 120µs + 4 × 10µs = 520µs
(c) With cut-through, the switch delays the packet by 200 bits = 2 µs. There
is still one 120 µs delay waiting for the last bit, and 20 µs of propagation
delay, so the total is 142 µs. To put it another way, the last bit still arrives
120 µs after the first bit; the first bit now faces two link delays and one
switch delay but never has to wait for the last bit along the way.
18. (a) The effective bandwidth is 100 Mbps; the sender can send data steadily
at this rate and the switches simply stream it along the pipeline. We are
assuming here that no ACKs are sent, and that the switches can keep up
and can buffer at least one packet.
(b) The data packet takes 520 µs as in 16(b) above to be delivered; the 400 bit
ACKs take 4 µs/link to be sent back, plus propagation, for a total of 4×4 µs
+4 × 10 µs = 56 µs; thus the total RTT is 576 µs. 12000 bits in 576 µs is
about 20.8 Mbps.
(c) 100×4.7×109
bytes / 12 hours = 4.7×1011
bytes/(12×3600s) ≈ 10.9 MBps
= 87 Mbps.
19. (a) 100×106
bps × 10 × 10−6
sec = 1000 bits = 125 bytes.

Chapter 1 4
(b) The first-bit delay is 520 µs through the store-and-forward switch, as in
16(a). 100 × 106
bps × 520×10−6
sec = 52000 bits = 650 bytes.
(c) 1.5×106
bps × 50 × 10−3
sec = 75,000 bits = 9375 bytes.
(d) The path is through a satellite, i.e. between two ground stations, not to
a satellite; this ground-to-satellite-to-ground path makes the total one-way
travel distance 2×35,900,000 meters. With a propagation speed of c =
3×108
meters/sec, the one-way propagation delay is thus 2×35,900,000/c
= 0.24 sec. Bandwidth×delay is thus 1.5 × 106
bps × 0.24 sec = 360,000
bits ≈ 45 KBytes
20. (a) Per-link transmit delay is 104
bits / 108
bps = 100 µs. Total transmission
time including link and switch propagation delays = 2×100+2×20+35 =
275 µs.
(b) When sending as two packets, the time to transmit one packet is cut in half.
Here is a table of times for various events:
T=0 start
T=50 A finishes sending packet 1, starts packet 2
T=70 packet 1 finishes arriving at S
T=105 packet 1 departs for B
T=100 A finishes sending packet 2
T=155 packet 2 departs for B
T=175 bit 1 of packet 2 arrives at B
T=225 last bit of packet 2 arrives at B
This is smaller than the answer to part (a) because packet 1 starts to make
its way through the switch while packet 2 is still being transmitted on the
first link, effectively getting a 50 µs head start. Smaller is faster, here.
21. (a) Without compression the total time is 1 MB/bandwidth. When we com-
press the file, the total time is
compression time + compressed size/bandwidth
Equating these and rearranging, we get
bandwidth = compression size reduction/compression time
= 0.5 MB/1 sec = 0.5 MB/sec for the first case,
= 0.6 MB/2 sec = 0.3 MB/sec for the second case.
(b) Latency doesn’t affect the answer because it would affect the compressed
and uncompressed transmission equally.
22. The number of packets needed, N, is ⌈106
/D⌉, where D is the packet data size.
Given that overhead = 50×N and loss = D (we have already counted the lost
packet’s header in the overhead), we have overhead+loss = 50 × ⌈106
/D⌉ + D.
D overhead+loss
1000 51000
10000 15000
20000 22500

Chapter 1 5
The optimal size is 10,000 bytes which minimizes the above function.
23. Comparison of circuits and packets result as follows :
(a) Circuits pay an up-front penalty of 1024 bytes being sent on one round trip
for a total data count of 2048 + n, whereas packets pay an ongoing per
packet cost of 24 bytes for a total count of 1024 × n/1000. So the question
really asks how many packet headers does it take to exceed 2048 bytes,
which is 86. Thus for files 86,000 bytes or longer, using packets results in
more total data sent on the wire.
(b) The total transfer latency for packets is the sum of the transmit delays,
where the per-packet transmit time t is the packet size over the bandwidth b
(8192/b), introduced by each of s switches (s × t), total propagation delay
for the links ((s + 2) × 0.002), the per packet processing delays introduced
by each switch (s×0.001), and the transmit delay for all the packets, where
the total packet count c is n/1000, at the source (c × t). Resulting in a
total latency of (8192s/b) + 0.003s + 0.004 + (8.192n/b) = (0.02924 +
0.000002048n) seconds. The total latency for circuits is the transmit delay
for the whole file (8n/b), the total propagation delay for the links, and the
setup cost for the circuit which is just like sending one packet each way
on the path. Solving the resulting inequality 0.02924 + 8.192(n/b) >
0.076576 + 8(n/b) for n shows that circuits achieve a lower delay for files
larger than or equal to 987,000B.
(c) Only the payload to overhead ratio size effects the number of bits sent,
and there the relationship is simple. The following table show the latency
results of varying the parameters by solving for the n where circuits become
faster, as above. This table does not show how rapidly the performance
diverges; for varying p it can be significant.
s b p pivotal n
5 4 Mbps 1000 987000
6 4 Mbps 1000 1133000
7 4 Mbps 1000 1280000
8 4 Mbps 1000 1427000
9 4 Mbps 1000 1574000
10 4 Mbps 1000 1721000
5 1 Mbps 1000 471000
5 2 Mbps 1000 643000
5 8 Mbps 1000 1674000
5 16 Mbps 1000 3049000
5 4 Mbps 512 24000
5 4 Mbps 768 72000
5 4 Mbps 1014 2400000
(d) Many responses are probably reasonable here. The model only considers
the network implications, and does not take into account usage of process-
ing or state storage capabilities on the switches. The model also ignores
the presence of other traffic or of more complicated topologies.

Chapter 1 6
24. The time to send one 12000-bit packet is 12000bits/100 Mbps = 120 µs. The
length of cable needed to exactly contain such a packet is 120 µs × 2×108
m/sec
= 24,000 meters.
12000 bits in 24000 meters is 50 bits per 100 m. With an extra 10 bits of delay in
each 100 m, we have a total of 60 bits/100 m or 0.6 bits/m. A 12000-bit packet
now fills 12000/(.6bits/m) = 20,000 meters.
25. For music we would need considerably more bandwidth, but we could toler-
ate high (but bounded) delays. We could not necessarily tolerate higher jitter,
though; see Section 6.5.1.
We might accept an audible error in voice traffic every few seconds; we might
reasonably want the error rate during music transmission to be a hundredfold
smaller. Audible errors would come either from outright packet loss, or from
jitter (a packet’s not arriving on time).
Latency requirements for music, however, might be much lower; a several-
second delay would be inconsequential. Voice traffic has at least a tenfold faster
requirement here.
26. (a) 640 × 480 × 3 × 30 bytes/sec = 26.4 MB/sec
(b) 160 × 120 × 1 × 5 = 96,000 bytes/sec = 94KB/sec
(c) 650MB/75 min = 8.7 MB/min = 148 KB/sec
(d) 8 × 10 × 72 × 72 pixels = 414,720 bits = 51,840 bytes. At 14,400 bits/sec,
this would take 28.8 seconds (ignoring overhead for framing and acknowl-
edgments).
28. (a) A file server needs lots of peak bandwidth. Latency is relevant only if
it dominates bandwidth; jitter and average bandwidth are inconsequential.
No lost data is acceptable, but without real-time requirements we can sim-
ply retransmit lost data.
(b) A print server needs less bandwidth than a file server (unless images are
extremely large). We may be willing to accept higher latency than (a), also.
(c) A file server is a digital library of a sort, but in general the world wide web
gets along reasonably well with much less peak bandwidth than most file
servers provide.
(d) For instrument monitoring we don’t care about latency or jitter. If data were
continually generated, rather than bursty, we might be concerned mostly
with average bandwidth rather than peak, and if the data really were routine
we might just accept a certain fraction of loss.
(e) For voice we need guaranteed average bandwidth and bounds on latency
and jitter. Some lost data might be acceptable; e.g. resulting in minor
dropouts many seconds apart.

Chapter 1 7
(f) For video we are primarily concerned with average bandwidth. For the sim-
ple monitoring application here, relatively modest video of Exercise 26(b)
might suffice; we could even go to monochrome (1 bit/pixel), at which
point 160×120×5frames/sec requires 12KB/sec. We could tolerate multi-
second latency delays; the primary restriction is that if the monitoring re-
vealed a need for intervention then we still have time to act. Considerable
loss, even of entire frames, would be acceptable.
(g) Full-scale television requires massive bandwidth. Latency, however, could
be hours. Jitter would be limited only by our capacity to absorb the arrival-
time variations by buffering. Some loss would be acceptable, but large
losses would be visually annoying.
29. In STDM the offered timeslices are always the same length, and are wasted if
they are unused by the assigned station. The round-robin access mechanism
would generally give each station only as much time as it needed to transmit,
or none if the station had nothing to send, and so network utilization would be
expected to be much higher.
30. (a) In the absence of any packet losses or duplications, when we are expecting
the Nth packet we get the Nth packet, and so we can keep track of N
locally at the receiver.
(b) The scheme outlined here is the stop-and-wait algorithm of Section 2.5;
as is indicated there, a header with at least one bit of sequence number is
needed (to distinguish between receiving a new packet and a duplication of
the previous packet).
(c) With out-of-order delivery allowed, packets up to 1 minute apart must be
distinguishable via sequence number. Otherwise a very old packet might
arrive and be accepted as current. Sequence numbers would have to count
as high as
bandwidth × 1 minute /packet size
31. In each case we assume the local clock starts at 1000.
(a) Latency: 100. Bandwidth: high enough to read the clock every 1 unit.
1000 1100
1001 1101
1002 1102
1003 1104 tiny bit of jitter: latency = 101
1004 1104
(b) Latency=100; bandwidth: only enough to read the clock every 10 units.
Arrival times fluctuate due to jitter.
1000 1100
1020 1110 latency = 90
1040 1145
1060 1180 latency = 120
1080 1184

Chapter 1 8
(c) Latency = 5; zero jitter here:
1000 1005
1001 1006
1003 1008 we lost 1002
1004 1009
1005 1010
32. Generally, with MAX PENDING =1, one or two connections will be accepted
and queued; that is, the data won’t be delivered to the server. The others will be
ignored; eventually they will time out.
When the first client exits, any queued connections are processed.
34. Note that UDP accepts a packet of data from any source at any time; TCP re-
quires an advance connection. Thus, two clients can now talk simultaneously;
their messages will be interleaved on the server.

1.
Clock
Bits 1 0 0 1 1 1 1 1 0 0 0 1 0 0 0 1
NRZ
Manchester
NRZI
2. See the figure below.
Bits 1 1 1 0 0 0 1 0 1 1 1 1 1 1 0 1 0 1 0 1
NRZI
4. One can list all 5-bit sequences and count, but here is another approach: there are
23
sequences that start with 00, and 23
that end with 00. There are two sequences,
00000 and 00100, that do both. Thus, the number that do either is 8+8−2 = 14,
and finally the number that do neither is 32 − 14 = 18. Thus there would have
been enough 5-bit codes meeting the stronger requirement; however, additional
codes are needed for control sequences.
5. The stuffed bits (zeros) are in bold:
1101 0111 1100 1011 1110 1010 1111 1011 0
6. The ∧ marks each position where a stuffed 0 bit was removed. There were no
stuffing errors detectable by the receiver; the only such error the receiver could
identify would be seven 1’s in a row.
1101 0111 11∧10 1111 1∧010 1111 1∧110
8. ..., DLE, DLE, DLE, ETX, ETX
9. (a) X DLE Y, where X can be anything besides DLE and Y can be anything
except DLE or ETX. In other words, each DLE must be followed by either
DLE or ETX.
(b) 0111 1111.
9

10. (a) After 48×8=384 bits we can be off by no more than ±1/2 bit, which is
about 1 part in 800.
(b) One frame is 810 bytes; at STS-1 51.8 Mbps speed we are sending 51.8×106
/(8×810)
= about 8000 frames/sec, or about 480,000 frames/minute. Thus, if station
B’s clock ran faster than station A’s by one part in 480,000, A would accu-
mulate about one extra frame per minute.
11. Suppose an undetectable three-bit error occurs. The three bad bits must be spread
among one, two, or three rows. If these bits occupy two or three rows, then some
row must have exactly one bad bit, which would be detected by the parity bit for
that row. But if the three bits are all in one row, then that row must again have a
parity error (as must each of the three columns containing the bad bits).
12. If we flip the bits corresponding to the corners of a rectangle in the 2-D layout of
the data, then all parity bits will still be correct. Furthermore, if four bits change
and no error is detected, then the bad bits must form a rectangle: in order for the
error to go undetected, each row and column must have no errors or exactly two
errors.
13. If we know only one bit is bad, then 2-D parity tells us which row and column it
is in, and we can then flip it. If, however, two bits are bad in the same row, then
the row parity remains correct, and all we can identify is the columns in which
the bad bits occur.
14. We need to show that the 1’s-complement sum of two non-0x0000 numbers
is non-0x0000. If no unsigned overflow occurs, then the sum is just the 2’s-
complement sum and can’t be 0000 without overflow; in the absence of overflow,
addition is monotonic. If overflow occurs, then the result is at least 0x0000 plus
the addition of a carry bit, i.e. ≥0x0001.
15. Let’s define swap([A,B]) = [B,A], where A and B are one byte each. We only
need to show [A, B] +’ [C, D] = swap([B, A] +’ [D, C]). If both (A+C) and
(B+D) have no carry, the equation obviously holds.
If A+C has a carry and B+D+1 does not,
[A, B] +’ [C, D] = [(A+C) & 0xEF, B+D+1]
swap([B, A] +’ [D, C]) = swap([B+D+1, (A+C) & 0xEF]) = [(A+C)
& 0xEF, B+D+1]
(The case where B+D+1 has also a carry is similar to the last case.)
If B+D has a carry, and A+C+1 does not,
[A, B] +’ [C, D] = [A+C+1, (B+D) & 0xEF].
swap([B, A] +’ [D, C]) = swap([(B+D) & 0xEF, A+C+1]) = [A+C+1,
(B+D) & 0xEF].
10

Chapter 2 11
If both (A+C) and (B+D) have a carry,
[A, B] +’ [C, D] = [((A+C) & 0xEF) + 1, ((B+D) & 0xEF) + 1]
swap([B, A] +’ [D, C]) = swap([((B+D) & 0xEF) + 1, ((A+C) &
0xEF) + 1] = [((A+C) & 0xEF) + 1, ((B+D) & 0xEF) + 1]
16. Consider only the 1’s complement sum of the 16-bit words. If we decrement a
low-order byte in the data, we decrement the sum by 1, and can incrementally
revise the old checksum by decrementing it by 1 as well. If we decrement a
high-order byte, we must decrement the old checksum by 256.
17. Here is a rather combinatorial approach. Let a, b, c, d be 16-bit words. Let [a, b]
denote the 32-bit concatenation of a and b, and let carry(a, b) denote the carry
bit (1 or 0) from the 2’s-complement sum a+b (denoted here a+2b). It suffices to
show that if we take the 32-bit 1’s complement sum of [a, b] and [c, d], and then
add upper and lower 16 bits, we get the 16-bit 1’s-complement sum of a, b, c,
and d. We note a +1 b = a +2 b +2 carry(a, b).
The basic case is supposed to work something like this. First,
[a, b] +2 [c, d] = [a +2 c +2 carry(b, d), b +2 d]
Adding in the carry bit, we get
[a, b] +1 [c, d] = [a +2 c +2 carry(b, d), b +2 d +2 carry(a, c)] (1)
Now we take the 1’s complement sum of the halves,
a +2 c +2 carry(b, d) +2 b +2 d +2 carry(a, c) + (carry(wholething))
and regroup:
= a +2 c +2 carry(a, c) +2 b +2 d +2 carry(b, d) + (carry(wholething))
= (a +1 c) +2 (b +1 d) + carry(a +1 c, b +1 d)
= (a +1 c) +1 (b +1 d)
which by associativity and commutativity is what we want.
There are a couple annoying special cases, however, in the preceding, where
a sum is 0xFFFF and so adding in a carry bit triggers an additional overflow.
Specifically, the carry(a, c) in (1) is actually carry(a, c, carry(b, d)), and sec-
ondly adding it to b +2 d may cause the lower half to overflow, and no provision
has been made to carry over into the upper half. However, as long as a +2 c and
b +2 d are not equal to 0xFFFF, adding 1 won’t affect the overflow bit and so the
above argument works. We handle the 0xFFFF cases separately.
Suppose that b +2 d = 0xFFFF =2 0. Then a +1 b +1 c +1 d = a +1 c. On the
other hand, [a, b] +1 [c, d] = [a +2 b, 0xFFFF] + carry(a, b). If carry(a, b) =
0, then adding upper and lower halves together gives a +2 b = a +1 b. If

Chapter 2 12
carry(a, b) = 1, we get [a, b] +1 [c, d] = [a +2 b +2 1, 0] and adding halves
again leads to a +1 b.
Now suppose a +2 c = 0xFFFF. If carry(b, d) = 1 then b +2 d 6= 0xFFFF
and we have [a, b] +1 [c, d] = [0, b +2 d +2 1] and folding gives b +1 d. The
carry(b, d) = 0 case is similar.
Alternatively, we may adopt a more algebraic approach. We may treat a buffer
consisting of n-bit blocks as a large number written in base 2n
. The numeric
value of this buffer is congruent mod (2n
−1) to the (exact) sum of the “digits”,
that is to the exact sum of the blocks. If this latter sum has more than n bits, we
can repeat the process. We end up with the n-bit 1’s-complement sum, which is
thus the remainder upon dividing the original number by 2n
− 1.
Let b be the value of the original buffer. The 32-bit checksum is thus b mod 232
−
1. If we fold the upper and lower halves, we get (b mod (232
−1)) mod (216
−1),
and, because 232
− 1 is divisible by 216
− 1, this is b mod (216
− 1), the 16-bit
checksum.
18. (a) We take the message 11100011, append 000 to it, and divide by 1001
according to the method shown in Section 2.4.3. The remainder is 100;
what we transmit is the original message with this remainder appended, or
1110 0011 100.
(b) Inverting the first bit of the transmission gives 0110 0011 100; dividing by
1001 (x3
+ 1) gives a remainder of 10; the fact that the remainder is non-
zero tells us a bit error occurred.
20. (b) p q C×q
000 000 000 000
001 001 001 101
010 011 010 111
011 010 011 010
100 111 100 011
101 110 101 110
110 100 110 100
111 101 111 001
(c) The bold entries 101 (in the dividend), 110 (in the quotient), and 101 110
in the body of the long division here correspond to the bold row of the
preceding table.

Chapter 2 13
110 101 011 100
1101 101 001 011 001 100
101 110
111 011
111 001
010 001
010 111
110 100
110 100
0
21. (a) M has eight elements; there are only four values for e, so there must be m1
and m2 in M with e(m1) = e(m2). Now if m1 is transmuted into m2 by a
two-bit error, then the error-code e cannot detect this.
(b) For a crude estimate, let M be the set of N-bit messages with four 1’s, and
all the rest zeros. The size of M is (N choose 4) = N!/(4!(N − 4)!).
Any element of M can be transmuted into any other by an 8-bit error. If
we take N large enough that the size of M is bigger than 232
, then as in
part (a) there must for any 32-bit error code function e(m) be elements
m1 and m2 of M with e(m1) = e(m2). To find a sufficiently large N,
we note N!/(4!(N − 4)!) > (N − 3)4
/24; it thus suffices to find N so
(N − 3)4
> 24 × 232
≈ 1011
. N ≈ 600 works. Considerably smaller
estimates are possible.
22. Assume a NAK is sent only when an out-of-order packet arrives. The receiver
must now maintain a RESEND NAK timer in case the NAK, or the packed it
NAK’ed, is lost.
Unfortunately, if the sender sends a packet and is then idle for a while, and this
packet is lost, the receiver has no way of noticing the loss. Either the sender
must maintain a timeout anyway, requiring ACKs, or else some zero-data filler
packets must be sent during idle times. Both are burdensome.
Finally, at the end of the transmission a strict NAK-only strategy would leave the
sender unsure about whether any packets got through. A final out-of-order filler
packet, however, might solve this.
23. (a) Propagation delay = 40 × 103
m/(2 × 108
m/s) = 200 µs.
(b) The roundtrip time would be about 400 µs. A plausible timeout time would
be twice this, or 0.8 ms. Smaller values (but larger than 0.4 ms!) might
be reasonable, depending on the amount of variation in actual RTTs. See
Section 5.2.6 of the text.
(c) The propagation-delay calculation does not consider processing delays that
may be introduced by the remote node; it may not be able to answer imme-
diately.
24. Bandwidth×(roundtrip)delay is about 125KBps × 2.5s = 312 KB, or 312 pack-
ets. The window size should be this large; the sequence number space must cover
twice this range, or up to 624. 10 bits are needed.

Chapter 2 14
26. If the receiver delays sending an ACK until buffer space is available, it risks de-
laying so long that the sender times out unnecessarily and retransmits the frame.
27. For Fig 2.17(b) (lost frame), there are no changes from the diagram in the text.
The next two figures correspond to the text’s Fig 2.17(c) and (d); (c) shows a
lost ACK and (d) shows an early timeout. For (c), the receiver timeout is shown
slightly greater than (for definiteness) twice the sender timeout.
Sender Receiver
Timeout
Timeout
Ignored
Timeout
Frame
Frame
ACK
ACK
Frame
Ignored
Sender Receiver
Timeout
duplicate frame; ignored;
receiver still waits for
timeout on Frame[N+1]
Frame[N]
Frame[N+1]
ACK[N]
Frame[N]
Timeout for Frame[N+1]
cancelled
ACK[N+1]
(c) (d)
Here is the version of Fig 2.17(c) (lost ACK), showing a receiver timeout of
approximately half the sender timeout.

Chapter 2 15
Sender Receiver
Frame[N]
Frame[N+1]
ACK
ACK
Timeout; receiver retransmits
before sender times out
Timeout
cancelled
ACK
Yet another Timeout
(possible, depending on
exact timeout intervals)
28. (a) The duplications below continue until the end of the transmission.
Sender Receiver
Frame[1]
ACK[2]
Frame[1]
ACK[2]
Frame[2]
Frame[2]
ACK[1]
ACK[1]
Frame[3]
Frame[3]
...
original ACK
response to duplicate frame
original frame
response to
duplicate ACK
original ACK
original frame
response to
duplicate ACK
original ACK
(b) To trigger the sorcerer’s apprentice phenomenon, a duplicate data frame
must cross somewhere in the network with the previous ACK for that frame.
If both sender and receiver adopt a resend-on-timeout strategy, with the
same timeout interval, and an ACK is lost, then both sender and receiver
will indeed retransmit at about the same time. Whether these retransmis-
sions are synchronized enough that they cross in the network depends on
other factors; it helps to have some modest latency delay or else slow hosts.
With the right conditions, however, the sorcerer’s apprentice phenomenon
can be reliably reproduced.
29. The following is based on what TCP actually does: every ACK might (optionally

Chapter 2 16
or not) contain a value the sender is to use as a maximum for SWS. If this value
is zero, the sender stops. A later ACK would then be sent with a nonzero SWS,
when a receive buffer becomes available. Some mechanism would need to be
provided to ensure that this later ACK is not lost, lest the sender wait forever. It
is best if each new ACK reduces SWS by no more than 1, so that the sender’s
LFS never decreases.
Assuming the protocol above, we might have something like this:
T=0 Sender sends Frame1-Frame4. In short order, ACK1...ACK4 are sent
setting SWS to 3, 2, 1, and 0 respectively.
The Sender now waits for SWS>0.
T=1 Receiver frees first buffer; sends ACK4/SWS=1.
Sender slides window forward and sends Frame5.
Receiver sends ACK5/SWS=0.
T=2 Receiver frees second buffer; sends ACK5/SWS=1.
Sender sends Frame6; receiver sends ACK6/SWS=0.
T=3 Receiver frees third buffer; sends ACK6/SWS=1.
T=4 Receiver frees fourth buffer; sends ACK7/SWS=1.
30. Here is one approach; variations are possible.
If frame[N] arrives, the receiver sends ACK[N] if NFE=N; otherwise if N was in
the receive window the receiver sends SACK[N].
The sender keeps a bucket of values of N>LAR for which SACK[N] was re-
ceived; note that whenever LAR slides forward this bucket will have to be purged
of all N≤LAR.
If the bucket contains one or two values, these could be attributed to out-of-
order delivery. However, the sender might reasonably assume that whenever
there was an N>LAR with frame[N] unacknowledged but with three, say, later
SACKs in the bucket, then frame[N] was lost. (The number three here is taken
from TCP with fast retransmit, which uses duplicate ACKs instead of SACKs.)
Retransmission of such frames might then be in order. (TCP’s fast-retransmit
strategy would only retransmit frame[LAR+1].)
31. The right diagram, for part (b), shows each of frames 4-6 timing out after a
2×RTT timeout interval; a more realistic implementation (e.g. TCP) would
probably revert to SWS=1 after losing packets, to address both congestion con-
trol and the lack of ACK clocking.

Chapter 2 17
Sender Receiver
Frame[1]
ACK[1]
Frame[2]
Frame[3]
ACK[2]
ACK[3]
Frame[4]
Frame[5]
Frame[6]
Frame[4] lost
Timeout Frame[4]
ACK[6]
Frame[7]
cumulative ACK
1 RTT
2 RTT
3 RTT
We might resend
ACK[3] here.
Sender Receiver
Frame[1]
ACK[1]
Frame[2]
Frame[3]
ACK[2]
ACK[3]
Frame[4]
Frame[5]
Frame[6]
Frames [4]-[6]
lost
Timeout
ACK[4]
Frame[7]
1 RTT
2 RTT
3 RTT
Frame[6]
Frame[5]
Frame[4]
Timeout
Timeout
ACK[5]
ACK[6]
...
33. In the following, ACK[N] means that all packets with sequence number less than
N have been received.
1. The sender sends DATA[0], DATA[1], DATA[2]. All arrive.
2. The receiver sends ACK[3] in response, but this is slow. The receive window
is now DATA[3]..DATA[5].
3. The sender times out and resends DATA[0], DATA[1], DATA[2]. For conve-
nience, assume DATA[1] and DATA[2] are lost. The receiver accepts DATA[0]
as DATA[5], because they have the same transmitted sequence number.
4. The sender finally receives ACK[3], and now sends DATA[3]-DATA[5]. The
receiver, however, believes DATA[5] has already been received, when DATA[0]
arrived, above, and throws DATA[5] away as a “duplicate”. The protocol now
continues to proceed normally, with one bad block in the received stream.
34. We first note that data below the sending window (that is, <LAR) is never sent
again, and hence – because out-of-order arrival is disallowed – if DATA[N] ar-
rives at the receiver then nothing at or before DATA[N-3] can arrive later. Simi-
larly, for ACKs, if ACK[N] arrives then (because ACKs are cumulative) no ACK

Chapter 2 18
before ACK[N] can arrive later. As before, we let ACK[N] denote the acknowl-
edgment of all data packets less than N.
(a) If DATA[6] is in the receive window, then the earliest that window can be
is DATA[4]-DATA[6]. This in turn implies ACK[4] was sent, and thus that
DATA[1]-DATA[3] were received, and thus that DATA[0], by our initial
remark, can no longer arrive.
(b) If ACK[6] may be sent, then the lowest the sending window can be is
DATA[3]..DATA[5]. This means that ACK[3] must have been received.
Once an ACK is received, no smaller ACK can ever be received later.
35. (a) The smallest working value for MaxSeqNum is 8. It suffices to show that
if DATA[8] is in the receive window, then DATA[0] can no longer arrive at
the receiver. We have that DATA[8] in receive window
⇒ the earliest possible receive window is DATA[6]..DATA[8]
⇒ ACK[6] has been received
⇒ DATA[5] was delivered.
But because SWS=5, all DATA[0]’s sent were sent before DATA[5]
⇒ by the no-out-of-orderarrival hypothesis, DATA[0] can no longer arrive.
(b) We show that if MaxSeqNum=7, then the receiver can be expecting DATA[7]
and an old DATA[0] can still arrive. Because 7 and 0 are indistinguishable
mod MaxSeqNum, the receiver cannot tell which actually arrived. We
follow the strategy of Exercise 27.
1. Sender sends DATA[0]...DATA[4]. All arrive.
2. Receiver sends ACK[5] in response, but it is slow. The receive window
is now DATA[5]..DATA[7].
3. Sender times out and retransmits DATA[0]. The receiver accepts it as
DATA[7].
(c) MaxSeqNum ≥ SWS + RWS.
36. (a) Note that this is the canonical SWS = bandwidth×delay case, with RTT =
4 sec. In the following we list the progress of one particular packet. At any
given instant, there are four packets outstanding in various states.
T=N Data[N] leaves A
T=N+1 Data[N] arrives at R
T=N+2 Data[N] arrives at B; ACK[N] leaves
T=N+3 ACK[N] arrives at R
T=N+4 ACK[N] arrives at A; DATA[N+4] leaves.
Here is a specific timeline showing all packets in progress:

Chapter 2 19
T=0 Data[0]...Data[3] ready; Data[0] sent
T=1 Data[0] arrives at R; Data[1] sent
T=2 Data[1] arrives at R; Data[0] arrives at B; ACK[0] starts back; Data[2] sent
T=3 ACK[0] arrives at R; Data[2] arrives at R; Data[1] arrives at B;
ACK[1] starts back; Data[3] sent
T=4 ACK[0] arrives at A; ACK[1] arrives at R; Data[3] arrives at R;
Data[2] arrives at B; ACK[2] starts back; Data[4] sent
T=5 ACK[1] arrives at A; ACK[2] arrives at R; Data[4] arrives at R;
Data[3] arrives at B; ACK[3] starts back; Data[5] sent
(b) T=0 Data[0]...Data[3] sent
T=1 Data[0]...Data[3] arrive at R
T=2 Data arrive at B; ACK[0]...ACK[3] start back
T=3 ACKs arrive at R
T=4 ACKs arrive at A; Data[4]...Data[7] sent
T=5 Data arrive at R
37. T=0 A sends frames 1-4. Frame[1] starts across the R–B link.
Frames 2,3,4 are in R’s queue.
T=1 Frame[1] arrives at B; ACK[1] starts back; Frame[2] leaves R.
Frames 3,4 are in R’s queue.
T=2 ACK[1] arrives at R and then A; A sends Frame[5] to R;
Frame[2] arrives at B; B sends ACK[2] to R.
R begins sending Frame[3]; frames 4,5 are in R’s queue.
Frame[3] arrives at B; B sends ACK[3] to R;
The steady-state queue size at R is two frames.
38. T=0 A sends frames 1-4. Frame[1] starts across the R–B link.
Frame[2] is in R’s queue; frames 3 & 4 are lost.
T=1 Frame[1] arrives at B; ACK[1] starts back; Frame[2] leaves R.
T=2 ACK[1] arrives at R and then A; A sends Frame[5] to R.
R immediately begins forwarding it to B.
T=3 ACK[2] arrives at R and then A; A sends Frame[6] to R.
R immediately begins forwarding it to B.
Frame[5] (not 3) arrives at B; B sends no ACK.
T=4 Frame[6] arrives at B; again, B sends no ACK.
T=5 A TIMES OUT, and retransmits frames 3 and 4.
R begins forwarding Frame[3] immediately, and enqueues 4.

Chapter 2 20
T=6 Frame[3] arrives at B and ACK[3] begins its way back.
R begins forwarding Frame[4].
T=7 Frame[4] arrives at B and ACK[6] begins its way back.
ACK[3] reaches A and A then sends Frame[7].
R begins forwarding Frame[7].
39. Hosts sharing the same address will be considered to be the same host by all
other hosts. Unless the conflicting hosts coordinate the activities of their higher
level protocols, it is likely that higher level protocol messages with otherwise
identical demux information from both hosts will be interleaved and result in
communication breakdown.
40. One-way delays:
Coax: 1500m 6.49 µs
link: 1000m 5.13 µs
repeaters two 1.20 µs
transceivers six 1.20 µs
(two for each repeater,
one for each station)
drop cable 6×50m 1.54 µs
Total: 15.56 µs
The roundtrip delay is thus about 31.1 µs, or 311 bits. The “official” total is
464 bits, which when extended by 48 bits of jam signal exactly accounts for the
512-bit minimum packet size.
The 1982 Digital-Intel-Xerox specification presents a delay budget (page 62 of
that document)that totals 463.8 bit-times, leaving 20 nanoseconds for unforeseen
contingencies.
41. A station must not only detect a remote signal, but for collision detection it must
detect a remote signal while it itself is transmitting. This requires much higher
remote-signal intensity.
42. (a) Assuming 48 bits of jam signal was still used, the minimum packet size
would be 4640+48 bits = 586 bytes.
(b) This packet size is considerably larger than many higher-level packet sizes,
resulting in considerable wasted bandwidth.
(c) The minimum packet size could be smaller if maximum collision domain
diameter were reduced, and if sundry other tolerances were tightened up.
43. (a) A can choose kA=0 or 1; B can choose kB=0,1,2,3. A wins outright if
(kA, kB) is among (0,1), (0,2), (0,3), (1,2), (1,3); there is a 5/8 chance of
this.
(b) Now we have kB among 0...7. If kA=0, there are 7 choices for kB that
have A win; if kA=1 then there are 6 choices. All told the probability of
A’s winning outright is 13/16.

Chapter 2 21
(c) P(winning race 1) = 5/8>1/2 and P(winning race 2) = 13/16>3/4; general-
izing, we assume the odds of A winning the ith race exceed (1 − 1/2i−1
).
We now have that
P(A wins every race given that it wins races 1-3)
≥ (1 − 1/8)(1 − 1/16)(1 − 1/32)(1 − 1/64)....
≈ 3/4.
(d) B gives up on it, and starts over with B2.
44. (a) If A succeeds in sending a packet, B will get the next chance. If A and B
are the only hosts contending for the channel, then even a wait of a fraction
of a slot time would be enough to ensure alternation.
(b) Let A and B and C be contending for a chance to transmit. We suppose the
following: A wins the first race, and so for the second race it defers to B
and C for two slot times. B and C collide initially; we suppose B wins the
channel from C one slot time later (when A is still deferring). When B now
finishes its transmission we have the third race for the channel. B defers
for this race; let us suppose A wins. Similarly, A defers for the fourth race,
but B wins.
At this point, the backoff range for C is quite high; A and B however are
each quickly successful – typically on their second attempt – and so their
backoff ranges remain bounded by one or two slot times. As each defers to
the other for this amount of time after a successful transmission, there is a
strong probability that if we get to this point they will continue to alternate
until C finally gives up.
(c) We might increase the backoff range given a decaying average of A’s recent
success rate.
45. If the hosts are not perfectly synchronized the preamble of the colliding packet
will interrupt clock recovery.
46. Here is one possible solution; many, of course, are possible. The probability of
four collisions appears to be quite low. Events are listed in order of occurrence.
A attempts to transmit; discovers line is busy and waits.
B attempts to transmit; discovers line is busy and waits.
C attempts to transmit; discovers line is busy and waits.
D finishes; A, B, and C all detect this, and attempt to transmit, and collide.
A chooses kA=1, B chooses kB=1, and C chooses kC=1.
One slot time later A, B, and C all attempt to retransmit, and again collide.
A chooses kA=2, B chooses kB=3, and C chooses kC=1.
One slot time later C attempts to transmit, and succeeds. While it transmits,
A and B both attempt to retransmit but discover the line is busy and wait.
C finishes; A and B attempt to retransmit and a third collision occurs. A
and B back off and (since we require a fourth collision) once again happen
to choose the same k < 8.

Chapter 2 22
A and B collide for the fourth time; this time A chooses kA=15 and B
chooses kB=14.
14 slot times later, B transmits. While B is transmitting, A attempts to
transmit but sees the line is busy, and waits for B to finish.
47. Many variations are, of course, possible. The scenario below attempts to demon-
strate several plausible combinations.
D finishes transmitting.
First slot afterwards: all three defer (P=8/27).
Second slot afterwards: A,B attempt to transmit (and collide); C defers.
Third slot: C transmits (A and B are presumably backing off, although no
relationship between p-persistence and backoff strategy was described).
C finishes.
First slot afterwards: B attempts to transmit and A defers, so B succeeds.
B finishes.
First slot afterwards: A defers.
Second slot afterwards: A defers.
Third slot afterwards: A defers.
Fourth slot afterwards: A defers a fourth time (P=16/81 ≈ 20%).
Fifth slot afterwards: A transmits.
A finishes.
48. (a) The second address must be distinct from the first, the third from the first
two, and so on; the probability that none of the address choices from
the second to the one thousand and twenty-fourth collides with an earlier
choice is
(1 − 1/248
)(1 − 2/248
) · · · (1 − 1023/248
)
≈ 1 − (1 + 2 + ... + 1023)/248
= 1 − 1, 047, 552/(2 × 248
).
Probability of a collision is thus 1, 047, 552/(2 ×248
) ≈ 1.86 ×10−9
. The
denominator should probably be 246
rather than 248
, since two bits in an
Ethernet address are fixed.
(b) Probability of the above on 220
≈ 1 million tries is 1.77 × 10−3
.
(c) Using the method of (a) yields (230
)2
/(2 × 248
) = 211
; we are clearly
beyond the valid range of the approximation. A better approximation, us-
ing logs, is presented in Exercise 8.18. Suffice it to say that a collision is
essentially certain.
49. (a) Here is a sample run. The bold backoff-time binary digits were chosen by
coin toss, with heads=1 and tails=0. Backoff times are then converted to
decimal.

Chapter 2 23
T=0: hosts A,B,C,D,E all transmit and collide. Backoff times are chosen
by a single coin flip; we happened to get kA=1, kB=0, kC=0, kD=1, kE=1.
At the end of this first collision, T is now 1. B and C retransmit at T=1; the
others wait until T=2.
T=1: hosts B and C transmit, immediately after the end of the first collision,
and collide again. This time two coin flips are needed for each backoff; we
happened to get kB = 00 = 0, kC = 11 = 3. At this point T is now 2; B
will thus attempt again at T=2+0=2; C will attempt again at T=2+3=5.
T=2: hosts A,B,D,E attempt. B chooses a three-bit backoff time as it is
on its third collision, while the others choose two-bit times. We got kA =
10 = 2, kB = 010 = 2, kD = 01 = 1, kE = 11 = 3. We add each k to
T=3 to get the respective retransmission-attempt times: T=5,5,4,6.
T=3: Nothing happens.
T=4: Station D is the only one to attempt transmission; it successfully
seizes the channel.
T=5: Stations A, B, and C sense the channel before transmission, but find
it busy. E joins them at T=6.
(b) Perhaps the most significant difference on a real Ethernet is that stations
close to each other will detect collisions almost immediately; only stations
at extreme opposite points will need a full slot time to detect a collision.
Suppose stations A and B are close, and C is far away. All transmit at the
same time T=0. Then A and B will effectively start their backoff at T≈0; C
will on the other hand wait for T=1. If A, B, and C choose the same backoff
time, A and B will be nearly a full slot ahead.
Interframe spacing is only one-fifth of a slot time and applies to all partici-
pants equally; it is not likely to matter here.
50. Here is a simple program:
#define USAGE "ether N"
// Simulates N ethernet stations all trying to transmit at once;
// returns average # of slot times until one station succeeds.
#include <iostream.h>
#include <stdlib.h>
#include <assert.h>
#define MAX 1000 /* max # of stations */
class station {
public:
void reset() { NextAttempt = CollisionCount = 0;}
bool transmits(int T) {return NextAttempt==T;}
void collide() { // updates station after a collision
CollisionCount ++;
NextAttempt += 1 + backoff( CollisionCount);

Chapter 2 24
//the 1 above is for the current slot
}
private:
int NextAttempt;
int CollisionCount;
static int backoff(int k) {
//choose random number 0..2∧
k-1; ie choose k random bits
unsigned short r = rand();
unsigned short mask = 0xFFFF >> (16-k); // mask = 2∧
k-1
return int (r & mask);
}
};
station S[MAX];
// run does a single simulation
// it returns the time at which some entrant transmits
int run (int N) {
int time = 0;
int i;
for (i=0;i<N;i++) { S[i].reset(); }
while(1) {
int count = 0; // # of attempts at this time
int j= -1; // count the # of attempts; save j as index of one of them
for (i=0; i<N; i++) {
if (S[i].transmits(time)) {j=i; ++count;}
}
if (count==1) // we are done
return time;
else if (count>1) { // collisions occurred
for (i=0;i<N;i++) {
if (S[i].transmits(time)) S[i].collide();
}
}
++time;
}
}
int RUNCOUNT = 10000;
void main(int argc, char * argv[]) {
int N, i, runsum=0;
assert(argc == 2);
N=atoi(argv[1]);
assert(N<MAX);
for (i=0;i<RUNCOUNT;i++) runsum += run(N);

Chapter 2 25
cout << "runsum = " << runsum
<< " RUNCOUNT= " << RUNCOUNT
<< " average: " << ((double)runsum)/RUNCOUNT << endl;
return;
}
Here is some data obtained from it:
# stations slot times
5 3.9
10 6.6
20 11.2
40 18.8
100 37.7
200 68.6
51. We alternate N/2 slots of wasted bandwidth with 5 slots of useful bandwidth.
The useful fraction is: 5/(N/2 + 5) = 10/(N+10)
52. (a) The program is below. It produced the following output:
λ # slot times λ # slot times
1 6.39577 2 4.41884
1.1 5.78198 2.1 4.46704
1.2 5.36019 2.2 4.4593
1.3 5.05141 2.3 4.47471
1.4 4.84586 2.4 4.49953
1.5 4.69534 2.5 4.57311
1.6 4.58546 2.6 4.6123
1.7 4.50339 2.7 4.64568
1.8 4.45381 2.8 4.71836
1.9 4.43297 2.9 4.75893
2 4.41884 3 4.83325
The minimum occurs at about λ=2; the theoretical value of the minimum
is 2e − 1 = 4.43656.
(b) If the contention period has length C, then the useful fraction is 8/(C +8),
which is about 64% for C = 2e − 1.
#include <iostream.h>
#include <stdlib.h>
#include <math.h>
const int RUNCOUNT = 100000;
// X = X(lambda) is our random variable
double X(double lambda) {
double u;
do {

Chapter 2 26
u= double(rand())/RAND MAX;
} while (u== 0);
double val = - log(u)*lambda;
return val;
}
double run(double lambda) {
int i = 0;
double time = 0;
double prevtime = -1;
double nexttime = 0;
time = X(lambda);
nexttime = time + X(lambda);
// while collision: adjacent times within +/- 1 slot
while (time - prevtime < 1 || nexttime - time < 1) {
prevtime = time;
time = nexttime;
nexttime += X(lambda);
}
return time;
}
void main(int argc, char * argv[]) {
int i;
double sum, lambda;
for (lambda = 1.0; lambda <= 3.01; lambda += 0.1) {
sum = 0;
for (i=0; i<RUNCOUNT; i++) sum += run(lambda);
cout << lambda << " " << sum/RUNCOUNT << endl;
}
}
53. This is the case in the hidden node problem, illustrated in Figure 2.30, in which
A interferes with C’s communication to B, and C interferes with A’s communi-
cation to B.
54. Whereas in wired networks, the sender can detect collisions as they occur, this
is not true in wireless. The hidden node problem (see previous question) is one
reason for this, and the half-duplex nature of wireless links is another.
55. 802.11 uses the RTS-CTS mechanism to try to address hidden terminals. A node
that has data to send begins by sending a short RTS packet indicating that it
would like to send data, and the receiver responds with a CTS, which is also
likely to be received by nodes that are in reach of the receiver but hidden from
the sender. While this doesn’t prevent collisions, the fact that RTS and CTS are
short packets makes collisions less likely.

Chapter 2 27
56. A base station topology would require an infrastructure of base stations in place.
Existing base stations (and their hardline connections) may be wiped out by the
disaster, and installing new ones may be difficult and time-consuming. With a
mesh topology, each additional node would piggyback on existing nodes.
57. GPS is considered too expensive and consumes too much power for the majority
of nodes. A typical solution requires a few nodes called beacons to determine
their own absolute locations based on GPS or manual configuration. The ma-
jority of nodes can then derive their absolute location by combining an estimate
of their position relative to the beacons with the absolute location information
provided by the beacons.

1. The following table is cumulative; at each part the VCI tables consist of the
entries at that part and also all previous entries. Note that at the last stage when
a connection comes in to A, we assume the VCI used at stage (a) cannot be re-
used in the opposite direction, as would be the case for bi-directional circuits
(the most common sort).
Exercise Switch Input Output
part Port VCI Port VCI
(a) 1 2 0 3 0
(b) 1 0 0 1 0
2 3 0 0 0
3 0 0 3 0
(c) 1 0 1 1 1
2 3 1 0 1
3 0 1 2 0
(d) 1 2 1 1 2
2 3 2 0 2
3 0 2 3 1
(e) 2 1 0 0 3
3 0 3 1 0
4 2 0 3 0
(f) 1 1 3 2 2
2 1 1 3 3
4 0 0 3 1
3. Node A: Destination Next hop Node B: Destination Next hop
B C A E
C C C E
D C D E
E C E E
F C F E
Node C: Destination Next hop Node D: Destination Next hop
A A A E
B E B E
D E C E
E E E E
F F F E
29

Chapter 3 30
Node E: Destination Next hop Node F: Destination Next hop
A C A C
B B B C
C C C C
D D D C
F C E C
4. S1: destination port
A 1
B 2
default 3
S2: destination port
A 1
B 1
C 3
D 3
default 2
C 2
D 3
default 1
D 2
default 1
5. In the following, Si[j] represents the jth entry (counting from 1 at the top) for
switch Si.
A connects to D via S1[1]—S2[1]—S3[1]
A connects to B via S1[2]
B connects to D via S1[3]—S2[2]—S3[2]
6. We provide space in the packet header for a second address list, in which we build
the return address. Each time the packet traverses a switch, the switch must add
the inbound port number to this return-address list, in addition to forwarding the
packet out the outbound port listed in the “forward” address. For example, as
the packet traverses Switch 1 in Figure 3.7, towards forward address “port 1”,
the switch writes “port 2” into the return address. Similarly, Switch 2 must write
“port 3” in the next position of the return address. The return address is complete
once the packet arrives at its destination.
Another possible solution is to assign each switch a locally unique name; that
is, a name not shared by any of its directly connected neighbors. Forwarding
switches (or the originating host) would then fill in the sequence of these names.
When a packet was sent in reverse, switches would use these names to look up
the previous hop. We might reject locally unique names, however, on the grounds

Chapter 3 31
that if interconnections can be added later it is hard to see how to permanently
allocate such names without requiring global uniqueness.
Note that switches cannot figure out the reverse route from the far end, given just
the original forward address. The problem is that multiple senders might use the
same forward address to reach a given destination; no reversal mechanism could
determine to which sender the response was to be delivered. As an example,
suppose Host A connects to port 0 of Switch 1, Host B connects to port 0 of
Switch 2, and Host C connects to port 0 of Switch 3. Furthermore, suppose port
1 of Switch 1 connects to port 2 of Switch 3, and port 1 of Switch 2 connects
to port 3 of Switch 3. The source-routing path from A to C and from B to C is
(0,1); the reverse path from C to A is (0,2) and from C to B is (0,3).
7. Here is a proposal that entails separate actions for (a) the switch that lost state,
(b) its immediate neighbors, and (c) everyone else. We will assume connec-
tions are bidirectional in that if a packet comes in on hport1,VCI1i bound for
hport2,VCI2i, then a packet coming in on the latter is forwarded to the former.
Otherwise a reverse-lookup mechanism would need to be introduced.
(a). A switch that has lost its state might send an I am lost message on its out-
bound links.
(b). Immediate neighbors who receive this would identify the port through which
the lost switch is reached, and then search their tables for any connection entries
that use this port. A connection broken message would be sent out the other port
of the connection entry, containing that port’s corresponding VCI.
(c). The remaining switches would then forward these connection broken mes-
sages back to the sender, forwarding them the usual way and updating the VCI
on each link.
A switch might not be aware that it has lost some or all of its state; one clue
is that it receives a packet for which it was clearly expected to have state, but
doesn’t. Such a situation could, of course, also result from a neighbor’s error.
8. If a switch loses its tables, it could notify its neighbors, but we have no means of
identifying what hosts down the line might use that switch.
So, the best we can do is notify senders by sending them an unable to forward
message whenever a packet comes in to the affected switch.
9. We now need to keep a network address along with each outbound port (or with
every port, if connections are bidirectional).
10. (a) The packet would be sent S1→S2→S3, the known route towards B. S3
would then send the packet back to S1 along the new connection, thinking
it had forwarded it to B. The packet would continue to circulate.
(b) This time it is the setup message itself that circulates forever.
11. As one possible answer, we use a modified version of Figure 3.44, in which hosts
H and J are removed, and port 0 of Switch 4 is connected to port 1 of Switch 3.

Chapter 3 32
Here are the hport,VCIi entries for a path from Host E to host F that traverses the
Switch2—Switch4 link twice; the VCI is 0 wherever possible.
Switch 2: h2,0i to h1,0i
Switch 4: h3,0i to h0,0i (recall Switch 4 port 0 now connects to Switch 3)
12. There is no guarantee that data sent along the circuit won’t catch up to and pass
the process establishing the connections, so, yes, data should not be sent until
the path is complete.
13.
15. When A sends to C, all bridges see the packet and learn where A is. However,
when C then sends to A, the packet is routed directly to A and B4 does not learn
where C is. Similarly, when D sends to C, the packet is routed by B2 towards B3
only, and B1 does not learn where D is.
B1: A-interface: A B2-interface: C (not D)
B2: B1-interface: A B3-interface: C B4-interface: D
B3: B2-interface: A,D C-interface: C
B4: B2-interface: A (not C) D-interface: D

Chapter 3 33
17. (a) When X sends to W the packet is forwarded on all links; all bridges learn
where X is. Y’s network interface would see this packet.
(b) When Z sends to X, all bridges already know where X is, so each bridge
forwards the packet only on the link towards X, that is, B3→B2→B1→X.
Since the packet traverses all bridges, all bridges learn where Z is. Y’s
network interface would not see the packet as B2 would only forward it on
the B1 link.
(c) When Y sends to X, B2 would forward the packet to B1, which in turn
forwards it to X. Bridges B2 and B1 thus learn where Y is. B3 and Z never
see the packet.
(d) When W sends to Y, B3 does not know where Y is, and so retransmits on
all links; Z’s network interface would thus see the packet. When the packet
arrives at B2, though, it is retransmitted only to Y (and not to B1) as B2
does know where Y is from step (c). B3 and B2 now know where W is, but
B1 does not learn where W is.
18. B1 will be the root; B2 and B3 each have two equal length paths (along their
upward link and along their downward link) to B1. They will each, indepen-
dently, select one of these vertical links to use (perhaps preferring the interface
by which they first heard from B1), and disable the other. There are thus four
possible solutions.
19. (a) The packet will circle endlessly, in both the M→B2→L→B1 and M→B1→
L→B2 directions.
(b) Initially we (potentially) have four packets: one from M clockwise, one
from M counterclockwise, and a similar pair from L.
Suppose a packet from L arrives at an interface to a bridge Bi, followed
immediately via the same interface by a packet from M. As the first packet
arrives, the bridge adds hL,arrival-interfacei to the table (or, more likely,
updates an existing entry for L). When the second packet arrives, addressed
to L, the bridge then decides not to forward it, because it arrived from the
interface recorded in the table as pointing towards the destination, and so it
dies.
Because of this, we expect that in the long run only one of the pair of pack-
ets traveling in the same direction will survive. We may end up with two
from M, two from L, or one from M and one from L. A specific scenario
for the latter is as follows, where the bridges’ interfaces are denoted “top”
and “bottom”:
1. L sends to B1 and B2; both place hL,topi in their table. B1 already has
the packet from M in the queue for the top interface; B2 this packet in the
queue for the bottom.
2. B1 sends the packet from M to B2 via the top interface. Since the
destination is L and hL,topi is in B2’s table, it is dropped.
3. B2 sends the packet from M to B1 via the bottom interface, so B1
updates its table entry for M to hM,bottomi

Chapter 3 34
4. B2 sends the packet from L to B1 via the bottom interface, causing it to
be dropped.
The packet from M now circulates counterclockwise, while the packet from
L circulates clockwise.
20. (a) In this case the packet would never be forwarded; as it arrived from a given
interface the bridge would first record hM,interfacei in its table and then
conclude the packet destined for M did not have to be forwarded out the
other interface.
(b) Initially we would have a copy of the packet circling clockwise (CW) and
a copy circling counterclockwise (CCW). This would continue as long as
they traveled in perfect symmetry, with each bridge seeing alternating ar-
rivals of the packet through the top and bottom interfaces. Eventually, how-
ever, something like the following is likely to happen:
0. Initially, B1 and B2 are ready to send to each other via the top interface;
both believe M is in the direction of the bottom interface.
1. B1 starts to send to B2 via the top interface (CW); the packet is somehow
delayed in the outbound queue.
2. B2 does send to B1 via the top interface (CCW).
3. B1 receives the CCW packet from step 2, and immediately forwards it
over the bottom interface back to B2. The CW packet has not yet been
delivered to B2.
4. B2 receives the packet from step 3, via the bottom interface. Because
B2 currently believes that the destination, M, lies on the bottom interface,
B2 drops the packet.
The clockwise packet would then be dropped on its next circuit, leaving the
loop idle.
21. (a) If the bridge forwards all spanning-tree messages, then the remaining bridges
would see networks D,E,F,G,H as a single network. The tree produced
would have B2 as root, and would disable the following links:
from B5 to A (the D side of B5 has a direct connection to B2)
from B7 to B
from B6 to either side
(b) If B1 simply drops the messages, then as far as the spanning-tree algorithm
is concerned the five networks D-H have no direct connection, and in fact
the entire extended LAN is partitioned into two disjoint pieces A-F and G-
H. Neither piece has any redundancy alone, so the separate spanning trees
that would be created would leave all links active. Since bridge B1 still
presumably is forwarding other messages, all the original loops would still
exist.
22. (a) Whenever any host transmits, the packet collides with itself.
(b) It is difficult or impossible to send status packets, since they too would self-
collide as in (a). Repeaters do not look at a packet before forwarding, so
they wouldn’t be in a position to recognize status packets as such.

Chapter 3 35
(c) A hub might notice a loop because collisions always occur, whenever any
host transmits. Having noticed this, the hub might send a specific signal
out one interface, during the rare idle moment, and see if that signal arrives
back via another. The hub might, for example, attempt to verify that when-
ever a signal went out port 1, then a signal always appeared immediately
at, say, port 3.
We now wait some random time, to avoid the situation where a neighboring
hub has also noticed the loop and is also disabling ports, and if the situation
still persists we disable one of the looping ports.
Another approach altogether might be to introduce some distinctive signal
that does not correspond to the start of any packet, and use this for hub-to-
hub communication.
23. Once we determine that two ports are on the same LAN, we can choose the
smaller-numbered port and shut off the other.
A bridge will know it has two interfaces on the same LAN when it sends out its
initial “I am root” configuration messages and receives its own messages back,
without their being marked as having passed through another bridge.
24. A 53-byte ATM cell has 5 bytes of headers, for an overhead of about 9.4% for
ATM headers alone. ATM adaptation layers add varying amounts of additional
overhead.
25. The drawbacks to datagram routing for small cells are the larger addresses, which
would now take up a considerable fraction of each cell, and the considerably
higher per-cell processing costs in each router that are not proportional to cell
size.
26. Since the I/O bus speed is less than the memory bandwidth, it is the bottle-
neck. Effective bandwidth that the I/O bus can provide is 800/2 Mbps because
each packet crosses the I/O bus twice. Therefore, the number of interfaces is
⌊400/100⌋ = 4.
28. The workstation can handle 1000/2 = 500 Mbps, limited by the I/O bus. Let the
packet size be x bits; to support 500,000 packets/second we need a total capacity
of 500000 × x bps; equating 5 × 105
× x = 500 × 106
bps, we get x = 1000
bits = 125 bytes. For packet sizes below this, packet forward rate is the limiter,
above this the limit is the I/O bus bandwidth.
29. Switch with input FIFO buffering :
(a) An input FIFO may become full if the packet at the head is destined for a
full output FIFO. Packets that arrive on ports whose input FIFOs are full
are lost regardless of their destination.
(b) This is called head-of-line blocking.

Chapter 3 36
(c) By redistributing the buffers exclusively to the output FIFOs, incoming
packets will only be lost if the destination FIFO is full.
30. Each stage has n/2 switching elements. Since after each stage we eliminate half
the network, i.e. half the rows in the n × n network, we need log2 n stages.
Therefore the number of switching elements needed is (n/2) log2 n. For n = 8,
this is 12.
31. A Batcher network sorts elements into ascending (or descending) order. As long
as no two elements have the same index, the ordered list can then be routed to
the correct output by a Banyan network. However, some additional mechanism
is needed to ensure that there are no duplicates. The paper by Giacopelli et al.
[GHMS91] gives one approach.
32. (a) After the upgrade the server—switch link is the only congested link. For a
busy Ethernet the contention interval is roughly proportional to the number
of stations contending, and this has now been reduced to two. So perfor-
mance should increase, but only slightly.
(b) A switch makes it impossible for a station to eavesdrop on traffic not ad-
dressed to it. On the other hand, switches tend to cost more than hubs, per
port.
33. IP addresses include the network/subnet, so that interfaces on different networks
must have different network portions of the address. Alternatively, addresses
include location information and different interfaces are at different locations,
topologically.
Point-to-point interfaces can be assigned a duplicate address (or no address) be-
cause the other endpoint of the link doesn’t use the address to reach the interface;
it just sends. Such interfaces, however, cannot be addressed by any other host
in the network. See also RFC1812, section 2.2.7, page 25, on “unnumbered
point-to-point links”.
34. The IPv4 header allocates only 13 bits to the Offset field, but a packet’s length
can be up to 216
− 1. In order to support fragmentation of a maximum-sized
packet, we must count offsets in multiples of 216−13
= 23
bytes.
The only concerns with counting fragmentation offsets in 8-byte units are that we
would waste space on a network with MTU = 8n + 7 bytes, or that alignment
on 8-byte boundaries would prove inconvenient. 8-byte chunks are small enough
that neither of these is a significant concern.
35. All 0’s or 1’s over the entire packet will change the Version and HLen fields,
resulting in non-IPv4 packets. The checksum algorithm would also catch this
error — a header consisting of all zeroes should have a checksum of all ones,
and vice versa. In reality, the checksum calculation would probably not even be
attempted after the version was found to be wrong.

Chapter 3 37
36. Consider the first network. An MTU of 1024 means that is the largest IP data-
gram that can be carried, so a datagram has room for 1024 − 20 = 1004 bytes
of IP-level data; because 1004 is not a multiple of 8, each fragment can contain
at most 8 × ⌊1004/8⌋ = 1000 bytes. We need to transfer 1024 + 20 = 1044
bytes of data when the TCP header is included. This would be fragmented into
fragments of size 1000, and 44.
Over the second network the 44-byte packet would be unfragmented but the
1000-data-byte packet would be fragmented as follows. The 576-byte MTU al-
lows for up to 576 − 20 = 556 bytes of payload, so rounding down to a multiple
of 8 again allows for 552 byes in the first fragment with the remaining 448 in the
second fragment.
38. (a) The probability of losing both transmissions of the packet would be 0.1×0.1
= 0.01.
(b) The probability of loss is now the probability that for some pair of identical
fragments, both are lost. For any particular fragment the probability of
losing both instances is 0.01 × 0.01 = 10−4
, and the probability that this
happens at least once for the 10 different fragments is thus about 10 times
this, or 0.001.
(c) An implementation might (though generally most do not) use the same
value for Ident when a packet had to be retransmitted. If the retransmis-
sion timeout was less than the reassembly timeout, this might mean that
case (b) applied and that a received packet might contain fragments from
each transmission.
39. M offset bytes data source
1 0 360 1st original fragment
1 360 152 1st original fragment
1 512 360 2nd original fragment
1 872 152 2nd original fragment
1 1024 360 3rd original fragment
0 1384 16 3rd original fragment
If fragmentation had been done originally for this MTU, there would be four
fragments. The first three would have 360 bytes each; the last would have 320
bytes.
40. The Ident field is 16 bits, so we can send 576 × 216
bytes per 60 seconds, or
about 5Mbps. If we send more than this, then fragments of one packet could
conceivably have the same Ident value as fragments of another packet.
41. IPv4 effectively requires that, if reassembly is to be done at the downstream
router, then it be done at the link layer, and will be transparent to IPv4. IP-layer
fragmentation is only done when such link-layer fragmentation isn’t practical,
in which case IP-layer reassembly might be expected to be even less practical,
given how busy routers tend to be. See RFC791, page 23.

Chapter 3 38
IPv6 uses link-layer fragmentation exclusively; experience had by then estab-
lished reasonable MTU values, and also illuminated the performance problems
of IPv4-style fragmentation. ( Path-MTU discovery is also mandatory, which
means the sender always knows just how large the data passed to IP can be to
avoid fragmentation.)
Whether or not link-layer fragmentation is feasible appears to depend on the
nature of the link; neither version of IP therefore requires it.
42. If the timeout value is too small, we clutter the network with unnecessary re-
requests, and halt transmission until the re-request is answered.
When a host’s Ethernet address changes, e.g. because of a card replacement,
then that host is unreachable to others that still have the old Ethernet address in
their ARP cache. 10-15 minutes is a plausible minimal amount of time required
to shut down a host, swap its Ethernet card, and reboot.
While self-ARP (described in the following exercise) is arguably a better solu-
tion to the problem of a too-long ARP timeout, coupled with having other hosts
update their caches whenever they see an ARP query from a host already in the
cache, these features were not always universally implemented. A reasonable
upper bound on the ARP cache timeout is thus necessary as a backup.
43. The answer is maybe, in theory, but the practical consequences rule it out. A
MAC address is statically assigned to each hardware interface. ARP mapping
enables indirection from IP addresses to the hardware MAC addresses. This
allows IP addresses to be dynamically reallocated when the hardware moves to
the different network, e.g. when a mobile wireless devices moves to a new access
network. So using MAC addresses as IP addresses would mean that we would
have to use static IP addresses.
Since the Internet routing takes advantage of address space hierarchy (use higher
bits for network addresses and lower bits for host addresses), if we would have
to use static IP addresses, the routing would be much less efficient. Therefore
this design is practically not feasible.
44. After B broadcasts any ARP query, all stations that had been sending to A’s phys-
ical address will switch to sending to B’s. A will see a sudden halt to all arriving
traffic. (To guard against this, A might monitor for ARP broadcasts purportedly
coming from itself; A might even immediately follow such broadcasts with its
own ARP broadcast in order to return its traffic to itself. It is not clear, however,
how often this is done.)
If B uses self-ARP on startup, it will receive a reply indicating that its IP address
is already in use, which is a clear indication that B should not continue on the
network until the issue is resolved.
45. (a) If multiple packets after the first arrive at the IP layer for outbound delivery,
but before the first ARP response comes back, then we send out multiple
unnecessary ARP packets. Not only do these consume bandwidth, but,

Chapter 3 39
because they are broadcast, they interrupt every host and propagate across
bridges.
(b) We should maintain a list of currently outstanding ARP queries. Before
sending a query, we first check this list. We also might now retransmit
queries on the list after a suitable timeout.
(c) This might, among other things, lead to frequent and excessive packet loss
at the beginning of new connections.
46. (a)
Information Distance to Reach Node
Stored at Node A B C D E F
A 0 ∞ 3 8 ∞ ∞
B ∞ 0 ∞ ∞ 2 ∞
C 3 ∞ 0 ∞ 1 6
D 8 ∞ ∞ 0 2 ∞
E ∞ 2 1 2 0 ∞
F ∞ ∞ 6 ∞ ∞ 0
(b)
A 0 ∞ 3 8 4 9
B ∞ 0 3 4 2 ∞
C 3 3 0 3 1 6
D 8 4 3 0 2 ∞
E 4 2 1 2 0 7
F 9 ∞ 6 ∞ 7 0
(c)
A 0 6 3 6 4 9
B 6 0 3 4 2 9
C 3 3 0 3 1 6
D 6 4 3 0 2 9
E 4 2 1 2 0 7
F 9 9 6 9 7 0

Chapter 3 40
48. D Confirmed Tentative
1. (D,0,-)
2. (D,0,-) (A,8,A)
(E,2,E)
3. (D,0,-) (A,8,A)
(E,2,E) (B,4,E)
(C,3,E)
4. (D,0,-) (A,6,E)
(E,2,E) (B,4,E)
(C,3,E) (F,9,E)
5. (D,0,-) (A,6,E)
(E,2,E) (F,9,E)
(C,3,E)
(B,4,E)
6. previous + (A,6,E)
7. previous + (F,9,E)
50. Traceroute sends packets with limited TTL values. If we send to an unassigned
network, then as long as the packets follow default routes, traceroute will get nor-
mal answers. When the packet reaches a default-free (backbone) router, however
(or more precisely a router which recognizes that the destination doesn’t exist),
the process will abruptly stop. Packets will not be forwarded further.
The router that finally realizes the error will send back “ICMP host unreachable”
or “ICMP net unreachable”, but this ICMP result may not in fact be listened for
by traceroute (is not, in implementations with which I am familiar), in which
case the traceroute session will end with timeouts either way.
51. A can reach B and D but not C. Because A hasn’t been configured with subnet
information, it treats C and B as being on the same network (it shares a network
number with them, being in the same site). To reach B, A sends ARP requests
directly to B; these are passed by RB as are the actual Ethernet packets. To reach
D, which A recognizes as being on another network, A uses ARP to send to R2.
However, if A tries to ARP to C, the request will not pass R1.
52. The cost=1 links show A connects to B and D; F connects to C and E.
F reaches B through C at cost 2, so B and C must connect.
F reaches D through E at cost 2, so D and E must connect.
A reaches E at cost 2 through B, so B and E must connect.
These give:

A

D

B

E

C

F
As this network is consistent with the tables, it is the unique minimal solution.

Chapter 3 41
54. (a) A: dest cost nexthop B: dest cost nexthop
B ∞ - A ∞ -
C 3 C C ∞ -
D ∞ - D 4 E
E ∞ - E 2 E
F 9 C F ∞ -
D: dest cost nexthop F: dest cost nexthop
A ∞ - A 9 C
B 4 E B ∞ -
C ∞ - C 6 C
E 2 E D ∞ -
F ∞ - E ∞ -
(b) A: dest cost nexthop D: dest cost nexthop
B 12 D A 8 A
C 3 C B 4 E
D 8 D C 11 A
E 10 D E 2 E
F 9 C F 17 A
(c) C: dest cost nexthop
A 3 A
B 15 A
D 11 A
E 13 A
F 6 F
55. Apply each subnet mask and if the corresponding subnet number matches the
SubnetNumber column, then use the entry in Next-Hop. (In these tables there is
always a unique match.)
(a) Applying the subnet mask 255.255.255.128, we get 128.96.39.0. Use in-
terface0 as the next hop.
(b) Applying subnet mask 255.255.255.128, we get 128.96.40.0. Use R2 as
the next hop.
(c) All subnet masks give 128.96.40.128 as the subnet number. Since there is
no match, use the default entry. Next hop is R4.
(d) Next hop is R3.
(e) None of the subnet number entries match, hence use default router R4.
57. (a) A necessary and sufficient condition for the routing loop to form is that B
reports to A the networks B believes it can currently reach, after A discov-
ers the problem with the A—E link, but before A has communicated to B
that A no longer can reach E.

Chapter 3 42
(b) At the instant that A discovers the A—E failure, there is a 50% chance that
the next report will be B’s and a 50% chance that the next report will be
A’s. If it is A’s, the loop will not form; if it is B’s, it will.
(c) At the instant A discovers the A—E failure, let t be the time until B’s next
broadcast. t is equally likely to occur anywhere in the interval 0 ≤ t ≤ 60.
The event of a loop forming is the same as the event that B broadcasts first,
which is the event that t 1.0 sec; the probability of this is 1/60.
58. Denote the act of A’s sending an update to B about E by A⇒B. Any initial
number of B⇒C or C⇒B updates don’t change E entries. By split horizon,
B⇒A and C⇒A are disallowed. Since we have assumed A reports to B before
C, the first relevant report must be A⇒B. This makes C the sole believer in
reachability of E; C’s table entry for E remains (E,2,A).
At this point legal and relevant updates are A⇒C, C⇒B, and B⇒C; A⇔B ex-
changes don’t change E entries and C⇒A is disallowed by split horizon. If A⇒C
or B⇒C the loop formation is halted, so we require C⇒B. Now C’s table has
(E,2,A) and B’s has (E,3,C); we have two believers.
The relevant possibilities now are B⇒A, or A⇒C. If B⇒A, then A’s table has
(E,4,C) and the loop is complete. If A⇒C, then B becomes the sole believer.
The only relevant update at that point not putting an end to belief in E is B⇒A,
which then makes A a believer as well.
At this point, exchange A⇒C would then form the loop. On the other hand,
C⇒B would leave A the sole believer. As things progress, we could either
(a) form a loop at some point,
(b) eliminate all belief in E at some point, or
(c) have sole-believer status migrate around the loop, C→B→A→C→ · · ·,
alternating with the dual-believer situation.
59. (a) The book already explains how poison reverse is not needed when F-G
fails. When the A-E link fails, the following sequence (or something sim-
ilarly bad) may happen depending on the timing, whether or not poison
reverse is used.
i. A sends (E, inf) to B.
ii. C sends (E, 2) to B. This route is via A.
iii. A sends (E, inf) to C.
iv. B sends (E, 3) to A. This route is via C.
v. C sends (E, inf) to B.
vi. A sends (E, 4) to C. This route is via B.
vii. B sends (E, inf) to A.
viii. C sends (E, 5) to B. This route is via A.
ix. A sends (E, inf) to C.
x. B sends (E, 6) to A. The oscillation goes on and on like this.

Chapter 3 43
(b) Without poison reverse, A and B would send each other updates that simply
didn’t mention X; presumably (this does depend somewhat on implemen-
tation) this would mean that the false routes to X would sit there until they
eventually aged out. With poison reverse, such a loop would go away on
the first table update exchange.
(c) 1. B and A each send out announcements of their route to X via C to each
other.
2. C announces to A and B that it can no longer reach X; the announce-
ments of step 1 have not yet arrived.
3. B and A receive each others announcements from step 1, and adopt
them.
60. We will implement hold-down as follows: when an update record arrives that
indicates a destination is unreachable, all subsequent updates within some given
time interval are ignored and discarded.
Given this, then in the EAB network A ignores B’s reachability news for one
time interval, during which time A presumably reaches B with the correct un-
reachability information.
Unfortunately, in the EABD case, this also means A ignores the valid B–D–E
path. Suppose, in fact, that A reports its failure to B, D reports its valid path
to B, and then B reports to A, all in rapid succession. This new route will be
ignored.
One way to avoid delaying discovery of the B–D–E path is to keep the hold-
down time interval as short as possible, relying on triggered updates to spread
the unreachability news quickly.
Another approach to minimizing delay for new valid paths is to retain route infor-
mation received during the hold-down period, but not to use it. At the expiration
of the hold-down period, the sources of such information might be interrogated
to determine whether it remains valid. Otherwise we might have to wait not
only the hold-down interval but also wait until the next regular update in order
to receive the new route news.
61. We will also assume that each node increments its sequence number only when
there is some change in the state of its local links, not for timer expirations (“no
packets time out”).
The central point of this exercise is intended to be an illustration of the “bringing-
up-adjacencies” process: in restoring the connection between the left- and right-
hand networks, it is not sufficient simply to flood the information about the re-
stored link. The two halves have evolved separately, and full information must
be exchanged.
Given that each node increments its sequence number whenever it detects a
change in its links to its neighbors, at the instant before the B—F link is restored
the LSP data for each node is as follows:

Chapter 3 44
node seq# connects to
A 2 B,C,D
B 2 A,C
C 2 A,B,D
D 2 A,C
F 2 G
G 2 F,H
H 1 G
When the B–F link is restored, OSPF has B and F exchange their full databases
of all the LSPs they have seen with each other. Each then floods the other side’s
LSPs throughout its side of the now-rejoined network. These LSPs are as in the
rows of the table above, except that B and F now each have sequence numbers
of 3.
The initial sequence number of an OSPF node is actually −231
+ 1.
62. Step confirmed tentative
1 (A,0,-)
2 (A,0,-) (D,2,D) (B,5,B)
3 (A,0,-) (D,2,D) (B,4,D) (E,7,D)
4 (A,0,-) (D,2,D) (B,4,D) (E,6,D) (C,8,D)
5 (A,0,-) (D,2,D) (B,4,D) (E,6,D) (C,7,D)
6 (A,0,-) (D,2,D) (B,4,D) (E,6,D) (C,7,D)
64. (a) This could happen if the link changed state recently, and one of the two
LSP’s was old.
(b) If flooding is working properly, and if A and B do in fact agree on the state
of the link, then eventually (rather quickly) whichever of the two LSP’s
was old would be updated by the same sender’s newer version, and reports
from the two sides of C would again agree.
65. This exercise does not, alas, quite live up to its potential.
The central idea behind Ethernet bridges is that they learn new host locations
by examining ordinary data packets, and do not receive new-host notices from
other bridges. Thus the first part of the final sentence of the exercise effectively
removes from consideration a genuine bridge-style approach for routers. While
there are good reasons for this, outlined in the final paragraph below, a better way
to phrase this might be to ask why IP routers do not work like learning bridges,
or, even more basically, why bridges do not use vector-distance routing.
Furthermore, a consequence of the second half of the final sentence is that there
is no real difference in the cases (a) and (b) with bridge-style learning. Proper
configuration would prevent address-assignment inconsistencies in each, which
apparently had been the original concern.

Chapter 3 45
So we are left with a model of “bridge-style learning” in which routers learn
about each other through messages each sends periodically to other routers. This
is not terribly bridge-like. Moreover, it is not clear what it means for routers
to learn of each other by this method; if they are sending each other messages
then either they already know about each other or else some form of broadcast
is used. And broadcast runs into serious problems if there is a possibility of
loops. If routers are sending out messages that are just broadcast on directly
connected subnets, listing all the subnets they know about, and these messages
include distance information, then they are more-or-less doing vector-distance
routing. One routing approach that might qualify under the terms of the exercise
is if routers send out link-state-style periodic messages identifying their directly
connected networks, and that these are propagated by flooding.
The main reason that IP routers cannot easily learn new subnet locations by ex-
amination of data packets is that they would then have to fall back on network-
wide broadcast for delivery to unknown subnets. IP does indeed support a notion
of broadcast, but broadcast in the presence of loop topology (which IP must sup-
port) fails rather badly unless specific (shortest-path) routes to each individual
subnet are already known by the routers. And even if some alternative mech-
anism were provided to get routing started, path-length information would not
be present in data packets, so future broadcasting would remain loop-unsafe.
We note too that subnet routing requires that the routers learn the subnet masks,
which are also not present in data packets. Finally, bridges may prefer passive
learning simply because it avoids bridge-to-bridge compatibility issues.
66. If an IP packet addressed to a specific host A were inadvertently broadcast, and
all hosts on the subnet did forwarding, then A would be inundated with multiple
copies of the packet.
Other reasons for hosts’ not doing routing include the risk that misconfigured
hosts could interfere with routing, or might not have up-to-date tables, or might
not even participate in the same routing protocol that the real routers were using.
68. (a) Giving each department a single subnet, the nominal subnet sizes are 27
,
26
, 25
, 25
respectively; we obtain these by rounding up to the nearest power
of 2. For example, a subnet with 128 addresses is large enough to contain
75 hosts. A possible arrangement of subnet numbers is as follows. Subnet
numbers are in binary and represent an initial segment of the bits of the last
byte of the IP address; anything to the right of the / represents host bits. The
/ thus represents the subnet mask. Any individual bit can, by symmetry, be
flipped throughout; there are thus several possible bit assignments.
A 0/ one subnet bit, with value 0; seven host bits
B 10/
C 110/
D 111/

Chapter 3 46
The essential requirement is that any two distinct subnet numbers remain
distinct when the longer one is truncated to the length of the shorter.
(b) We have two choices: either assign multiple subnets to single departments,
or abandon subnets and buy a bridge. Here is a solution giving A two
subnets, of sizes 64 and 32; every other department gets a single subnet of
size the next highest power of 2:
A 01/
001/
B 10/
C 000/
D 11/
69. To solve this with routing, C has to be given its own subnet. Even if this is
small, this reduces the available size of the original Ethernet to at most seven
bits of subnet address. Here is a possible routing table for B; subnet numbers
and masks are in binary. Note that many addresses match neither subnet.
net subnet mask interface
200.0.0 0/000 0000 1000 0000 Ethernet
200.0.0 1000 00/00 1111 1100 direct link
Here C’s subnet has been made as small as possible; only two host bits are avail-
able (a single host bit can’t be used because all-zero-bits and all-ones-bits are re-
served in the host portion of an address). C’s address might now be 200.0.0.10000001,
with the last octet again in binary.
70. (a) A would broadcast an ARP request “where is C?”
B would answer it; it would supply its own Ethernet address.
A would send C’s packet to B’s Ethernet address.
B would forward the packet to C.
(b) For the above to work, B must know to forward the packet without using
subnet addressing; this is typically accomplished by having B’s routing
table contain a “host-specific route”:
net/host interface
C direct link
200.0.0 Ethernet
Host-specific routes must be checked first for this to work.
71. (a) DHCP will have considerable difficulty sorting out to which subnet var-
ious hosts belonged; subnet assignments would depend on which server
answered first. The full DHCP deals with this by allowing servers to be
manually configured to ignore address-assignment requests from certain
physical addresses on the other subnet. Note that subnet assignment in this
situation may matter for consistent naming, performance reasons, certain
security access rules, and for broadcast protocols.

Chapter 3 47
(b) ARP will not be affected. Hosts will only broadcast ARP queries for other
hosts on the same subnet; hosts on the other subnet will hear these but
won’t answer. A host on one subnet would answer an ARP query from the
other subnet, if it were ever issued, but it wouldn’t be.
72. (a): B (b): A (c): E (d): F (e): C (f): D
(For the last one, note that the first 14 bits of C4.6B and C4.68 match.)
74. (a) each department expects the growth in the number of machines as follows
• Engineering expects machine number increase by one per week, thus
by 52 per year. Note that we need 5 machines initially.
• Sales expects client number increase by (−1)·0.20 +(+1)·0.60 +0 ·
0.20 = 0.40, thus machine number increase by 0.40 · 1/2 = 0.20 per
week, so by 0.20 ∗ 52 = 10.4 per year. Note that we do not need any
machines in the first year, but at the beginning of the second year, we
need 3 machines since we have 6 clients then.
• Marketing expects no increase.
To guarantee addresses for at least seven years, we need (5 + 52 · 7) + (3 +
10.4 · 6) + 16 = 450.4 addresses. Therefore, the new company needs a
slash 23 address range to accommodate 512 addresses.
(b) To determine how long the 512 addresses allocation would last: suppose it
would last n years, (5+52·n)+(3+10.4·(n−1))+16 = 13.6+62.4·n =
512. Thus, n = 7.99. The address assignment would be, for engineering,
5 + 52 · n = 420.32 ∼ 420, for sales, (3 + 10.4 · (n − 1)) = 75.67 ∼ 76,
for marketing, 16.
(c) Since class B supports 65534 host addresses and class C supports 254 ad-
dresses (note that two addresses are always reserved in each network class),
the company could get one class B range or two class C ranges.
75. There are many possible answers to this problem. One approach might be to use
a tree of all netmasks. We decide on the left versus right subtree at level i based
on the ith bit of the address. A network with an n-bit mask is marked in the tree
at level n. Given an address, we use the bits to proceed down the tree until we
reach a dead end. At that point we use the last-encountered network; this ensures
the longest match was found. Such a tree is sometimes called a trie.
This strategy is linear in the address size. Performance might be enhanced by
handling 4 or even 8 address bits at each level of the tree, although this would
lead to some increased space used.
Another approach might be to maintain a separate dictionary for each n, 1 ≤
n ≤ 24, of all masks of length n. We start with the longest mask length and
work backwards, at stage n searching for the first n bits of the address in the
length-n dictionary. If dictionary lookup were sufficiently fast this might also be
roughly linear in address length.

1. (a) Q will receive three routes to P, along links 1, 2, and 3.
(b) A−→B traffic will take link 1. B−→A traffic will take link 2. Note that
this strategy minimizes cost to the source of the traffic.
(c) To have B−→A traffic take link 1, Q could simply be configured to prefer
link 1 in all cases. The only general solution, though, is for Q to accept into
its routing tables some of the internal structure of P, so that Q for example
knows where A is relative to links 1 and 2.
(d) If Q were configured to prefer AS paths through R, or to avoid AS paths
involving links 1 and 2, then Q might route to P via R.
2. In the diagram below, the shortest path between A and B (measured by number
of router hops) passes through AS P, AS Q, and AS P.
!#$#
!# %#
!##
While such a path might be desirable (the path via Q could be much faster or
offer lower latency, for example), BGP would see the same AS number (for AS
P) twice in the AS PATH. To BGP, such an AS PATH would appear as a loop,
and be disallowed.
3. (a) The diameter D of a network organized as a binary tree, with root node as
“backbone”, would be of order log2 A. The diameter of a planar rectangular
grid of connections would be of order
√
A.
(b) For each AS S, the BGP node needs to maintain a record of the AS PATH
to S, requiring 2×actual path length bytes. It also needs a list of all the net-
works within S, requiring 4×number of networks bytes. Summing these
up for all autonomous systems, we get 2AD + 4N, or 2AC log A + 4N
and 2AC
√
A + 4N for the models from part (a), where C is a constant.
4. Many arrangements are possible, although perhaps not likely. Here is an allo-
cation scheme that mixes two levels of geography with providers; it works with
48-bit InterfaceIDs. The subdivisions become much more plausible with 64-bit
InterfaceIDs.
48

Bytes 0-1: 3-bit prefix + country where site is located
(5 bits is not enough to specify the country)
Bytes 2-3: provider
Bytes 4-5: Geographical region within provider
Bytes 6-8: Subscriber (large providers may have 64K subscribers)
Bytes 8-9: (Byte 8 is oversubscribed) Subnet
Bytes 10-15: InterfaceID
5. (a) P’s table:
address nexthop
C2.0.0.0/8 Q
C3.0.0.0/8 R
C1.A3.0.0/16 PA
C1.B0.0.0/12 PB
Q’s table:
address nexthop
C1.0.0.0/8 P
C3.0.0.0/8 R
C2.0A.10.0/20 QA
C2.0B.0.0/16 QB
R’s table:
address nexthop
C1.0.0.0/8 P
C2.0.0.0/8 Q
(b) The same, except for the following changes of one entry each to P’s and
R’s tables:
P: C3.0.0.0/8 Q // was R
R: C1.0.0.0/8 Q // was P
(c) Note the use of the longest-match rule to distinguish the entries for Q
QA in P’s table, and for P PA in Q’s table.
P’s table:
address nexthop
C2.0.0.0/8 Q
C2.0A.10.0/20 QA // for QA
C1.A3.0.0/16 PA
C1.B0.0.0/12 PB
Q’s table:
address nexthop
C1.0.0.0/8 P
C1.A3.0.0/16 PA // for PA
C2.0A.10.0/20 QA
C2.0B.0.0/16 QB
6. The longest-match rule is intended for this. Note that all providers now have to
include entries for PA and QB, though.
49

P’s table:
address nexthop
C2.0.0.0/8 Q
C3.0.0.0/8 R
C1.A3.0.0/16 Q // entry for PA
C1.B0.0.0/12 PB
C2.0B.0.0/16 R // entry for QB
Q’s table:
address nexthop
C1.0.0.0/8 P
C3.0.0.0/8 R
C1.A3.0.0/16 PA // now Q’s customer
C2.0A.10.0/20 QA
C2.0B.0.0/16 R // entry for QB
R’s table:
address nexthop
C1.0.0.0/8 P
C2.0.0.0/8 Q
C1.A3.0.0/16 Q // R also needs an entry for PA
C2.0B.0.0/16 QB // QB is now R’s customer
7. (a) Inbound traffic takes a single path to the organization’s address block,
which corresponds to the organization’s “official” location. This means
all traffic enters the organization at a single point even if much shorter al-
ternative routes exist.
(b) For outbound traffic, the organization could enter into its own tables all the
highest-level geographical blocks for the outside world, allowing the orga-
nization to route traffic to the exit geographically closest to the destination.
(c) For an approach such as the preceding to work for inbound traffic as well,
the organization would have to be divided internally into geographically
based subnets, and the outside world would then have to accept routing
entries for each of these subnets. Consolidation of these subnets into a
single external entry would be lost.
(d) We now need each internal router to have entries for internal routes to all
the other internal IP networks; this suffices to ensure internal traffic never
leaves.
8. Perhaps the primary problem with geographical addressing is what to do with
geographically dispersed sites that have their own internal connections. Routing
all traffic to a single point of entry seems inefficient.
At the time when routing-table size became a critical issue, most providers were
regional and thus provider-based addressing was more or less geographical.
9. As described, ISP X is a customer of both ISP A and ISP B. If he advertises
a path learned from A to ISP B, then B may send him traffic that he will then
50

have to forward on to A. At the least, this consumes resources for ISP X without
producing any income, and it may even increase his costs if he pays either A or
B based on volume. Hence, it would not typically be a good idea for ISP X to
advertise such a path.
10. (a) If Q does not advertise A to the world, then only traffic originating within
Q will take the Q—A link. Other traffic will be routed first to P. If Q does
advertise A, then traffic originating at an external site B will travel in via Q
whenever the B–Q–A path is shorter than the B–P–A path.
(b) Q must advertise A’s reachability to the world, but it may put a very low
“preference value” on this link.
(c) The problem is that most outbound traffic will take the DEFAULT path, and
nominally this is a single entry. Some mechanism for load-sharing must be
put into place. Alternatively, A could enter into its internal routing tables
some of its most common external IP destinations, and route to these via
Q.
11. (a) R1 should be configured to forward traffic destined for outside of A to the
new ISP. R2 should route traffic for A to A as before. Note that if a host
on N sends its outbound traffic to R2 by mistake, R2 will send it via the
old link. R2 should continue to advertise N to the rest of A. N’s outbound
traffic would then take the new link, but inbound traffic would still travel
via A. Subnets are not announced into the backbone routing tables, so R1
would not announce N to the world.
(b) If N has its own IP network number, then R1 does announce its route to N
to the world. R1 would not necessarily announce its route to A, however.
R2 would not change: it would still announce N into A. Assuming A does
not announce its route to N into its provider, all external traffic to and from
N now goes through the new link, and N-A traffic goes through R2.
(c) If A wants to use N’s R1 link as a backup, then R1 needs to announce to
the backbone that it has a route to A, but give this route a cost higher than
that of A’s original link (techniques for doing this via BGP include route
preference indications and “padding” the report with extra ASs.)
12. IP has a subrange of its address space reserved for multicast addresses. In IPv4,
these addresses are assigned in the class D address space, and IPv6 also has a
portion of its address space (see Table 4.11) reserved for multicast group ad-
dresses.
13. Yes. Some subranges of the multicast ranges are reserved for intradomain mul-
ticast, so they can be reused independently by different domains.
14. The host must have joined at least one of the other 31 IP multicast groups whose
addresses share the same high-order 5 bits and hence the same Ethernet multicast
address.
51

16. See figure on following page. Note that we have assumed that all hosts except
the sources are members of G. (This was unclear in the first printing.)
17. (a) One multicast transmission involves all k + k2
+ ... + kN−1
= (kN
−
k)/(k − 1) links.
(b) One unicast retransmission involves N links; sending to everyone would
require N × kN
links.
(c) Unicast transmission to x fraction of the recipients uses x × N × kN
links.
Equating this to the answer in (a), we get x = (kN
− k)/((k − 1) × N ×
kN
) ≈ 1/(k − 1) × N
18. (a) In the PIM-SM scheme, each router in a multicast group independently
decides when to create a source-specific tree for a particular source. The
router does so by sending a source-specific Join message toward the source.
Hosts connected to the router benefit from this in the form of decreased
latency of the multicasts they receive. The intention is that the router would
do so only in response to a high data rate being observed from that source,
since it imposes a cost, in additional routing state, on other routers. An
unscrupulously configured router, however, could indiscriminately trigger
source-specific trees, without the justification of high data rates.
(b) In the PIM-SSM scheme, any host can join a source-specific group, thereby
triggering creation of a source-specific tree with the attendant source-specific
routing state. This is presumably not a problem since whoever assigned the
SSM address to the group (a subrange of the IP multicast addresses is re-
served for SSM) did so because SSM was appropriate for the particular
group.
19. A simple example where the routes are different has a connected triangle of
routers: the source router, the group member router, and a router on the RP
address’s link. A simple example where the routes are the same has the source
and group member routers each connected only to a router on the RP address’s
link.
20. (a) correct
(b) incorrect (::: is not defined as abbreviating notation)
(c) incorrect (shorthand can only be used for one set of contiguous 0’s)
(d) correct
(e) incorrect (an IPv4 address mapped to IPv6 should be preceded by FFFF
hex).
21. First, MPLS labels are of link-local scope—this means that the same label can
be used on different links to mean different things. This in turn means that the
number of labels needed on a link is just the number of forwarding equivalence
classes (FECs) that are meaningful on that link. Thus, if each label is used to
represent a prefix in the routing table, as described in Section 4.5.1, then up to a
million prefixes could be handled with 20 bits.
52

Source
Source
Figure 1: Answers to question 4.16
53

Chapter 4 54
22. MPLS has been thought to improve router performance because each label is
a direct index in the routing table, and thus an MPLS-only router could avoid
running the more complex longest IP prefix match algorithm. But packet for-
warding has many other aspects that influence performance, such as enqueueing
packets and switching them across a backplane. These aspects are independent
of the forwarding algorithm and have turned out to be the dominant performance-
influencing factors.
23. (a) 8 bytes are needed to attach two MPLS labels.
(b) 20 bytes are needed for an additional IP header.
(c) Bandwidth efficiency for MPLS is 300/308 = 0.97, and for IP is 300/320
= 0.94. For 64-byte packets, MPLS has 64/72 = 0.89 and IP has 64/84 =
0.76. MPLS is relatively more efficient when the payload size is smaller.
24. Source routing cannot specify a long path because of the option size limit. Sec-
ond, IP option processing is considerably more complex than normal IP forward-
ing, and can cause significant performance penalties. Finally, source routing can-
not readily aggregate the traffic with the same route into one class; by contrast,
MPLS can aggregate such traffic as one FEC, represented by a single label at
each hop, thus improving scalability.
25. A correspondent node has no way of knowing that the IP address of a mobile
node has changed, and hence no way to send it a packet. A TCP connection will
break if the IP address of one endpoint changes.
26. The home agent and the mobile node may be very far apart, leading to suboptimal
routing.
27. Without some sort of authentication of updates, an attacker could tell the corre-
spondent node to send all the traffic destined for a mobile node to a node that the
attacker controls, thus stealing the traffic. Or, an attacker can tell any number of
correspondent nodes to send traffic to some other node that the attacker wishes
to flood with traffic.

1. (a) An application such as TFTP, when sending initial connection requests,
might want to know the server isn’t accepting connections.
(b) On typical Unix systems, one needs to open a socket with attribute IP RAW
(traditionally requiring special privileges) and receive all ICMP traffic.
(c) A receiving application would have no way to identify ICMP messages as
such, or to distinguish between these messages and protocol-specific data.
2. (a) In the following, the client receives file “foo” when it thinks it has requested
“bar”.
1. The client sends a request for file “foo”, and immediately aborts locally.
The request, however, arrives at the server.
2. The client sends a new request, for file “bar”. It is lost.
3. The server responds with first data packet of “foo”, answering the only
request it has actually seen.
(b) Requiring the client to use a new port number for each separate request
would solve the problem. To do this, however, the client would have to
trust the underlying operating system to assign a new port number each
time a new socket was opened. Having the client attach a timestamp or
random number to the file request, to be echoed back in each data packet
from the server, would be another approach fully under the application’s
control.
3. The TFTP protocol is a reasonable model although with some idiosyncrasies that
address other issues; see RFC 1350. TFTP’s first packet, called Read Request,
RRQ, simply names a file. Upon receipt, the server creates a new ephemeral
port from which to answer, and begins sending data from that new port. The
client assumes that the first well-formed packet it receives is this server data, and
records the data’s source port. Any subsequent packets from a different port are
discarded and an error response is sent.
The basic stop-and-wait transfer is standard, although one must decide if se-
quence numbers are allowed to wrap around or not. Here are approaches, TFTP’s
and otherwise, for (a)-(c):
(a) The most basic approach here is to require the server to keep track of con-
nections, as long as they are active. The problem with this is that the client
is likely to be simply an application, and can exit at any time. It may exit
and retransmit a request for a different file, or a new request for the same
file, before the server knows there was a problem or status change.
A more robust mechanism for this situation might be a CONNECT NUM
field, either chosen randomly or clock-driven or incremented via some cen-
tral file for each client connection attempt. Such a field would correspond
roughly with TCP’s initial sequence number (ISN).
55

Chapter 5 56
In TFTP, if the RRQ is duplicated then the server might well create two
processes and two ports from which to answer. (A server that attempted to
do otherwise would have to maintain considerable state about past RRQ’s.)
Whichever process contacted the client first would win out, though, while
the other would receive an error response from the client. In one sense,
then, duplicate TFTP RRQ’s do duplicate the connection, but only one of
the duplicates survives.
(b) The TFTP approach here is to have the client enter a “dallying” period af-
ter the final data was received, so that the process is still around (perhaps
moved to the background) to receive and re-acknowledge any retransmis-
sions of the final data. This period roughly corresponds to TIME WAIT.
(c) The dallying approach of (b) also ties up the client socket for that period,
preventing another incarnation of the connection. (However, TFTP has no
requirement that dallying persist for a time interval approaching the MSL.)
TFTP also specifies that both sides are to choose “random” port numbers
for each connection (although “random” is generally interpreted as “as-
signed by the operating system”). If either side chooses a new port num-
ber, then late-arriving packets don’t interfere even if the other side reuses its
previous port number. A CONNECT NUM field would also be effective
here.
4. Host A has sent a FIN segment to host B, and has moved from ESTABLISHED
to FIN WAIT 1. Host A then receives a segment from B that contains both
the ACK of this FIN, and also B’s own FIN segment. This could happen if the
application on host B closed its end of the connection immediately when the host
A’s FIN segment arrived, and was thus able to send its own FIN along with the
ACK.
Normally, because the host B application must be scheduled to run before it can
close the connection and thus have the FIN sent, the ACK is sent before the FIN.
While “delayed ACKs” are a standard part of TCP, traditionally only ACKs of
DATA, not FIN, are delayed. See RFC 813 for further details.
5. The two-segment-lifetime timeout results from the need to purge old late dupli-
cates, and uncertainty of the sender of the last ACK as to whether it was received.
For the first issue we only need one connection endpoint in TIME WAIT; for the
second issue, a host in the LAST ACK state expects to receive the last ACK,
rather than send it.
6. The receiver includes the advertised window in the ACKs to the sender. The
sender probes the receiver to know when the advertised window becomes greater
than 0; if the receiver’s ACK advertising a larger window is lost, then a later
sender probe will elicit a duplicate of that ACK.
If responsibility for the lost window-size-change ACK is shifted from the sender
to the receiver, then the receiver would need a timer for managing retransmission
of this ACK until the receiver were able to verify it had been received.

Chapter 5 57
A more serious problem is that the receiver only gets confirmation that the sender
has received the ACK when new data arrives, so if the connection happens to fall
idle the receiver may be wasting its time.
8. The sequence number doesn’t always begin at 0 for a transfer, but is randomly
or clock generated.
9. (a) The advertised window should be large enough to keep the pipe full; de-
lay (RTT) × bandwidth here is 100 ms × 1 Gbps = 100 Mb = 12.5 MB of
data. This requires 24 bits if we assume the window is measured in bytes
(224
≈= 16million) for the AdvertisedWindow field. The sequence num-
ber field must not wrap around in the maximum segment lifetime. In 30
seconds, 30 Gb = 3.75 GB can be transmitted. 32 bits allows a sequence
space of about 4GB, and so will not wrap in 30 seconds. (If the maximum
segment lifetime were not an issue, the sequence number field would still
need to be large enough to support twice the maximum window size; see
“Finite Sequence Numbers and Sliding Window” in Section 2.5.)
(b) The bandwidth is straightforward from the hardware; the RTT is also a
precise measurement but will be affected by any future change in the size
of the network. The MSL is perhaps the least certain value, depending as it
does on such things as the size and complexity of the network, and on how
long it takes routing loops to be resolved.
11. The problem is that there is no way to determine whether a packet arrived on the
first attempt or whether it was lost and retransmitted.
Having the receiver echo back immediately and measuring the elapsed times
would help; many Berkeley-derived implementations measure timeouts with a
0.5 sec granularity and round-trip times for a single link without loss would gen-
erally be one to two orders of magnitude smaller. But verifying that one had such
an implementation is itself rather difficult.
12. (a) This is 125MB/sec; the sequence numbers wrap around when we send
232
B = 4 GB. This would take 4GB/(125MB/sec) = 32 seconds.
(b) Incrementing every 32 ms, it would take about 32 × 4 × 109
ms, or about
four years, for the timestamp field to wrap.
14. (a) If a SYN packet is simply a duplicate, its ISN value will be the same as
the initial ISN. If the SYN is not a duplicate, and ISN values are clock-
generated, then the second SYN’s ISN will be different.
(b) We will assume the receiver is single-homed; that is, has a unique IP ad-
dress. Let hraddr, rporti be the remote sender, and lport be the local port.
We suppose the existence of a table T indexed by hlport, raddr, rporti

Chapter 5 58
and containing (among other things) data fields lISN and rISN for the local
and remote ISNs.
if (connections to lport are not being accepted)
send RST
else if (there is no entry in T for hlport, raddr, rporti) // new SYN
Put hlport, raddr, rporti into a table,
Set rISN to be the received packet’s ISN,
Set lISN to be our own ISN,
Send the reply SYN+ACK
Record the connection as being in state SYN RECD
else if (T [hlport, raddr, rporti] already exists)
if (ISN in incoming packet matches rISN from the table)
// SYN is a duplicate; ignore it
else
send RST to hraddr, rporti)
15. x = y if and only if (y − x) ≥ 0, where the expression y − x is taken to be
signed even though x and y are not.
16. (a) A would send an ACK to B for the new data. When this arrived at B,
however, it would lie outside the range of “acceptable ACKs” and so B
would respond with its own current ACK. B’s ACK would be acceptable to
A, and so the exchanges would stop.
If B later sent less than 100 bytes of data, then this exchange would be
repeated.
(b) Each end would send an ACK for the new, forged data. However, when re-
ceived both these ACKs would lie outside the range of “acceptable ACKs”
at the other end, and so each of A and B would in turn generate their current
ACK in response. These would again be the ACKs for the forged data, and
these ACKs would again be out of range, and again the receivers would
generate the current ACKs in response. These exchanges would continue
indefinitely, until one of the ACKs was lost.
If A later sent 200 bytes of data to B, B would discard the first 100 bytes
as duplicate, and deliver to the application the second 100 bytes. It would
acknowledge the entire 200 bytes. This would be a valid ACK for A.
For more examples of this type of scenario, see Joncheray, L; A Simple Ac-
tive Attack Against TCP; Proceedings of the Fifth USENIX UNIX Security
Symposium, June, 1995.
17. Let H be the host to which A had been connected; we assumed B is able to guess
H. As we are also assuming telnet connections, B can restrict probes to H’s telnet
port (port 23).
First, B needs to find a port A had been using. For various likely ephemeral
port numbers N, B sends an ACK packet from port N to hH,telneti. For many
implementations, ephemeral ports start at some fixed value (e.g. N=1024) and
increase sequentially; for an unshared machine it is unlikely that very many ports

Chapter 5 59
had been used. If A had had no connection from port N, H will reply to B with a
RST packet. But if H had had an outstanding connection to hA,Ni, then H will
reply with either nothing (if B’s forged ACK happened to be Acceptable, i.e. in
the current window at the point when A was cut off), or the most recent Accept-
able ACK (otherwise). Zero-byte data packets can with most implementations
also be used as probes here.
Once B finds a successful port number, B then needs to find the sequence number
H is expecting; once B has this it can begin sending data on the connection as
if it were still A. To find the sequence number, B again takes advantage of the
TCP requirement that H reply with the current ACK if B sends an ACK or DATA
inconsistent with H’s current receive window [that is, an “unacceptable ACK”].
In the worst case B’s first probe lies in H’s window, in which case B needs to
send a second probe.
18. We keep a table T, indexed by haddress,porti pairs, and containing an integer
field for the ISN and a string field for the connection’s DATA.
We will use = for sequence number comparison as in Exercise 15.
if (SYN flag is set in P.TCPHEAD.Flags)
Create the entry T[hP.IPHEAD.SourceAddr,P.TCPHEAD.SrcPorti]
T[...].ISN = P.TCPHEAD.SequenceNum
T[...].DATA = hempty stringi
else
See if DATA bit in P.TCPHEAD.Flags is set; if not, ignore
Look up T[hP.IPHEAD.SourceAddr,P.TCPHEAD.SrcPorti]
(if not found, ignore the packet)
See if P.TCPHEAD.SequenceNum = T[...].ISN+100.
If so, append the appropriate portion of the packet’s data to T[...].DATA
19. (a) 1. C connects to A, and gets A’s current clock-based ISNA1.
2. C sends a SYN packet to A, purportedly from B. A sends SYN+ACK,
with ISNA2 to B, which we are assuming is ignored.
3. C makes a guess at ISNA2, e.g. ISNA1 plus some suitable increment,
and sends the appropriate ACK to A, along with some data that has some
possibly malign effect on A. As in Step 2, this packet too has a forged
source address of B.
4. C does nothing further, and the connection either remains half-open
indefinitely or else is reset, but the damage is done.
(b) In one 40 ms period there are 40 ms/4µsec = 10,000 possible ISNAs; we
would expect to need about 10,000 tries.
Further details can be found in Morris, RT; A Weakness in the 4.2BSD UNIX
TCP/IP Software; Computing Science Technical Report No. 117, ATT Bell
Laboratories, Murray Hill, NJ, 1985.

Chapter 5 60
20. (a) T=0.0 ‘a’ sent
T=1.0 ‘b’ collected in buffer
T=2.0 ‘c’ collected in buffer
T=3.0 ‘d’ collected in buffer
T=4.0 ‘e’ collected in buffer
T=4.1 ACK of ‘a’ arrives, “bcde” sent
T=5.0 ‘f’ collected in buffer
T=6.0 ‘g’ collected in buffer
T=7.0 ‘h’ collected in buffer
T=8.0 ‘i’ collected in buffer
T=8.2 ACK arrives; “fghi” sent
(b) The user would type ahead blindly at times. Characters would be echoed
between 4 and 8 seconds late, and echoing would come in chunks of four
or so. Such behavior is quite common over telnet connections, even those
with much more modest RTTs, but the extent to which this is due to the
Nagle algorithm is unclear.
(c) With the Nagle algorithm, the mouse would appear to skip from one spot
to another. Without the Nagle algorithm the mouse cursor would move
smoothly, but it would display some inertia: it would keep moving for one
RTT after the physical mouse were stopped. (We’ve assumed in this case
that the mouse and the display are at the same end of the connection.)
21. (a) We have 4096 ports; we eventually run out if the connection rate averages
more than 4096/60 = 70 per sec. (The range used here for ephemeral ports,
while small, is typical of older TCP implementations.)
(b) In the following we let A be the host that initiated the close (and that is in
TIME WAIT); the other host is B. A is nominally the client; B the server.
If B fails to receive an ACK of its final FIN, it will eventually retransmit
that FIN. So long as A remains in TIME WAIT it is supposed to reply
again with the corresponding ACK. If the sequence number of the FIN
were incorrect, A would send RST.
If we allow reopening before TIME WAIT expires, then a given very-late-
arriving FIN might have been part of any one of a number of previous
connections. For strict compliance, host A would have to maintain a list
of prior connections, and if an old FIN arrived (as is theoretically possible,
given that we are still within the TIME WAIT period for the old connec-
tion), host A would consult this list to determine whether the FIN had an
appropriate sequence number and hence whether an ACK or RST should
be sent.
Simply responding with an ACK to all FINs with sequence numbers before
the ISN of the current connection would seem reasonable, though. The
old connection, after all, no longer exists at B’s end to be reset, and A
knows this. A knows, in fact, that a prior final ACK or RST that it sent in
response to B’s FIN was received by B, since B allowed the connection to
be reopened, and so it might justifiably not send anything.

Chapter 5 61
22. Whichever endpoint remains in TIME WAIT must retain a record of the connec-
tion for the duration of TIME WAIT; as the server typically is involved in many
more connections than clients, the server’s record-keeping requirements would
be much more onerous.
Note also that some implementations of TIME WAIT simply disallow all new
connections to the port in question for the duration, not only those from the par-
ticular remote connection that initiated the TIME WAIT. Since a server cannot
choose a new port, this might mean it could process at most one connection per
TIME WAIT interval.
In situations where the client requests some variable-length stream (eg a file), the
server might plausibly initiate the active close to indicate the end of the data.
23. Timeouts indicates that the network is congested and that one should send fewer
packets rather than more. Exponential backoff immediately gives the network
twice as long to deliver packets (though a single linear backoff would give the
same); it also rapidly adjusts to even longer delays, thus it in theory readily ac-
commodating sharp increases in RTT without further loading the already over-
taxed routers. If the RTT suddenly jumps to 15 times the old TimeOut, expo-
nential increase retransmits at T=1, 3, 7, and 15; linear increase would retransmit
at T=1, 3, 6, 10, and 15. The difference here is not large. Exponential backoff
makes the most difference when the RTT has increased by a very large amount,
either due to congestion or network reconfiguration, or when “polling” the net-
work to find the initial RTT.
24. The probability that a Normally distributed random variable is more than π stan-
dard deviations above the mean is about 0.0816%.
25. If every other packet is lost, we transmit each packet twice.
(a) Let E ≥ 1 be the value for EstimatedRTT, and T = 2×E be the value for
TimeOut. We lose the first packet and back off TimeOut to 2 × T . Then,
when the packet arrives, we resume with EstimatedRTT = E, TimeOut =
T . In other words, TimeOut doesn’t change.
(b) Let T be the value for TimeOut. When when we transmit the packet the
first time, it will be lost and we will wait time T . At this point we back
off and retransmit using TimeOut = 2 × T . The retransmission succeeds
with an RTT of 1 sec, but we use the backed-off value of 2 × T for the next
TimeOut. In other words, TimeOut doubles with each received packet.
This is Not Good.
26. Using initial Deviation =1.0 it took 21 iterations for TimeOut to fall below 4.0.
With an initial Deviation of 0.1, it took 20 iterations; with an initial Deviation
of 2 it took 22.

Chapter 5 62
Iteration SampleRTT EstRTT Dev diff TimeOut
0 1.00 4.00 1.00
1 1.00 3.63 1.25 -3.00 8.63
2 1.00 3.31 1.42 -2.63 8.99
3 1.00 3.03 1.53 -2.31 9.15
4 1.00 2.78 1.59 -2.03 9.14
5 1.00 2.56 1.61 -1.78 9.00
6 1.00 2.37 1.61 -1.56 8.81
7 1.00 2.20 1.58 -1.37 8.52
8 1.00 2.05 1.54 -1.20 8.21
9 1.00 1.92 1.48 -1.05 7.84
10 1.00 1.81 1.41 -.92 7.45
11 1.00 1.71 1.34 -.81 7.07
12 1.00 1.63 1.27 -.71 6.71
13 1.00 1.56 1.19 -.63 6.32
14 1.00 1.49 1.12 -.56 5.97
15 1.00 1.43 1.05 -.49 5.63
16 1.00 1.38 .98 -.43 5.30
17 1.00 1.34 .91 -.38 4.98
18 1.00 1.30 .84 -.34 4.66
19 1.00 1.27 .78 -.30 4.39
20 1.00 1.24 .72 -.27 4.12
21 1.00 1.21 .66 -.24 3.85
28. One approach to this, shown below, is to continue the table above, except that
whenever TimeOut would fall below 4.0 we replace the SampleRTT of that
row with 4.0.
We could also create a table starting from scratch, using an initial Estimate-
dRTT of 1.0 and seeding the first few rows with a couple instances of Sam-
pleRTT = 4.0 to get TimeOut ≥ 4.0 in the first place.
Either way, N is between 6 and 7 here.

Chapter 5 63
row # SampleRTT EstRTT Dev diff TimeOut
19 1.00 1.24 0.72 -0.27 4.13
20 4.00 1.58 0.98 2.76 5.50
21 1.00 1.51 0.93 -0.58 5.22
22 1.00 1.45 0.88 -0.51 4.95
23 1.00 1.39 0.82 -0.45 4.68
24 1.00 1.34 0.77 -0.39 4.42
25 1.00 1.30 0.72 -0.34 4.16
26 4.00 1.64 0.96 2.70 5.49
27 1.00 1.56 0.92 -0.64 5.25
28 1.00 1.49 0.88 -0.56 4.99
29 1.00 1.43 0.83 -0.49 4.74
30 1.00 1.37 0.78 -0.43 4.48
31 1.00 1.33 0.73 -0.37 4.24
32 4.00 1.66 0.97 2.67 5.54
29. Here is the table of the updates to the EstRTT, etc statistics. Packet loss is ig-
nored; the SampleRTTs given may be assumed to be from successive singly
transmitted segments. Note that the first column, therefore, is simply a row num-
ber, not a packet number, as packets are sent without updating the statistics when
the measurements are ambiguous. Note also that both algorithms calculate the
same values for EstimatedRTT; only the TimeOut calculations vary.
new TimeOut old TimeOut
SampleRTT EstRTT Dev diff EstRTT+4×Dev 2×EstRTT
1.00 0.10 1.40 2.00
1 5.00 1.50 0.59 4.00 3.85 3.00
2 5.00 1.94 0.95 3.50 5.74 3.88
3 5.00 2.32 1.22 3.06 7.18 4.64
4 5.00 2.66 1.40 2.68 8.25 5.32
New algorithm (TimeOut = EstimatedRTT+ 4×Deviation):
There are a total of three retransmissions, two for packet 1 and one for packet 3.
The first packet after the change times out at T=1.40, the value of TimeOut at
that moment. It is retransmitted, with TimeOut backed off to 2.8. It times out
again 4.2 sec after the first transmission, and TimeOut is backed off to 5.6.
At T=5.0 the first ACK arrives and the second packet is sent, using the backed-off
TimeOut value of 5.6. This second packet does not time out, so this constitutes
an unambiguous RTT measurement, and so timing statistics are updated to those
of row 1 above.
When the third packet is sent, with TimeOut=3.85, it times out and is retrans-
mitted. When its ACK arrives the fourth packet is sent, with the backed-off
TimeOut value, 2×3.85 = 7.70; the resulting RTT measurement is unambiguous
so timing statistics are updated to row 2. When the fifth packet is sent, Time-
Out=5.74 and no further timeouts occur.

Chapter 5 64
If we continue the above table to row 9, we get the maximum value for TimeOut,
of 10.1, at which point TimeOut decreases toward 5.0.
Original algorithm (TimeOut = 2×EstimatedRTT):
There are five retransmissions: for packets 1, 2, 4, 6, 8.
The first packet times out at T=2.0, and is retransmitted. The ACK arrives before
the second timeout, which would have been at T=6.0.
When the second packet is sent, the backed-off TimeOut of 4.0 is used and we
time out again. TimeOut is now backed off to 8.0. When the third packet is sent,
it thus does not time out; statistics are updated to those of row 1.
The fourth packet is sent with TimeOut=3.0. We time out once, and then trans-
mit the fifth packet without timeout. Statistics are then updated to row 2.
This pattern continues. The sixth packet is sent with TimeOut = 3.88; we
again time out once, send the seventh packet without loss, and update to row
3. The eighth packet is sent with TimeOut=4.64; we time out, back off, send
packet 9, and update to row 4. Finally the tenth packet does not time out, as
TimeOut=2×2.66=5.32 is larger than 5.0.
TimeOut continues to increase monotonically towards 10.0, as EstimatedRTT
converges on 5.0.
30. Let the real RTT (for successful transmissions) be 1.0 units. By hypothesis, ev-
ery packet times out once and then the retransmission is acknowledged after 1.0
units; this means that each SampleRTT measurement is TimeOut+1 = Esti-
matedRTT+1. We then have
EstimatedRTT = α× EstimatedRTT + β× SampleRTT
= EstimatedRTT + β×(SampleRTT −EstimatedRTT).
≥ EstimatedRTT + β
Thus it follows that the Nth EstimatedRTT is greater than or equal to Nβ.
Without the assumption TimeOut = EstimatedRTT we still have SampleRTT
− EstimatedRTT ≥ 1 and so the above argument still applies.
31. For the steady state, assume the true RTT is 3 and EstimatedRTT is 1. At T=0
we send a data packet. Since TimeOut is twice EstimatedRTT=1, at T=2 the
packet is retransmitted. At T=3 the ACK of the original packet returns (because
the true RTT is 3); measured SampleRTT is thus 3−2 = 1; this equals Estimat-
edRTT and so there is no change. This is illustrated by the following diagram:

Chapter 5 65
Sender Receiver
Packet
ACK Packet
EstimatedRTT
SampleRTT
T=0
T=1
T=2
T=3
Timeout and
Retransmission
To get to such a steady state, assume that originally RTT = EstimatedRTT =
1.45, say, and RTT then jumps to 3.0 as above. The first packet sent under the
new rules will time out and be retransmitted at T=2.9; when the ACK arrives at
T=3.0 we record SampleRTT = 0.1. This causes EstimatedRTT to decrease.
It will continue to grow smaller, monotonically (at least if β is not too large),
converging on the value 1.0 as above.
32. A FIN or RST must lie in the current receive window. A RST outside this win-
dow is ignored; TCP responds to an out-of-window FIN with the current ACK:
If an incoming segment is not acceptable, an acknowledgment should
be sent in reply (unless the RST bit is set, if so drop the segment and
return) [RFC793]
Note that a RST can lie anywhere within the current window; its sequence num-
ber need not be the next one in sequence.
If a FIN lies in the current window, then TCP waits for any remaining data and
closes the connection. If a RST lies in the current window then the connection is
immediately closed:
If the RST bit is set then, any outstanding RECEIVEs and SEND
should receive “reset” responses. All segment queues should be flushed.
Users should also receive an unsolicited general “connection reset”
signal. Enter the CLOSED state, delete the TCB, and return.
33. (a) The first incarnation of the connection must have closed successfully and
the second must have opened; this implies the exchange of FIN and SYN
packets and associated ACKs. The delayed data must also have been suc-
cessfully retransmitted.
(b) One plausible hypothesis is that two routes were available from one host
to the other. Traffic flowed on one link, which suddenly developed severe
congestion or routing-loop delays at which point subsequent traffic was
switched to the other, now-faster, link. It doesn’t matter whether the two

Chapter 5 66
routes were both used initially on an alternating basis, or if the second route
was used only after the first route failed.
34. We suppose A is connected to B and wishes to hand off the connection to C.
There is more than one approach; in the following we assume that A and C do
most of the work, and that A decides on a point in the sequence number stream
after which B is to send to C. It could be argued that B is better positioned
to make that determination. C also continues here with the sequence numbers
started by A.
New function call event: handoff(); allowed in ESTABLISHED state only.
New packet types:
HANDOFF REQ // request from A to C
HANDOFF DO // request from A to B
New states:
HANDOFF CALLED // for A
H REQ SENT // for A, HANDOFF REQ sent to C
H REQ ACK // for A; C has acknowledged HANDOFF REQ
H REQ RECD // for C
H START WAIT // for C
H TIME WAIT // for A
Here is a chronology of events.
1. handoff() called. A moves to state HANDOFF CALLED, and identifies a
sequence number H SEQ (for B) after which data is to be sent to C. A waits
for B to send up to this sequence number, blocking further transmissions by
shrinking the upper edge of the receive window to H SEQ. Whether or not A
buffers data following H SEQ, and forwards it to C, is optional.
2. A sends HANDOFF REQ to C, with sequence number H SEQ−1 (from B)
and A’s own current sequence number. C moves to state H REQ RECD. A
moves to state H REQ SENT. If A has been buffering data past H SEQ, it might
send it to C at this point.
3. C sends an ACK to A to accept the handoff (or RST to reject it). If the former,
A moves to state H REQ ACK. C moves to H START WAIT, and waits to hear
from B.
4. A sends HANDOFF DO to B, with H SEQ. B remains ESTABLISHED, and
sends an ACK to A, which moves to H TIME WAIT. B also sends an ACK to C,
which moves to ESTABLISHED.
Any data with sequence number before H SEQ that arrives at A during the
H TIME WAIT period is now forwarded by A to C.
35. (a) In order to disallow simultaneous open, an endpoint in state SYN SENT
should not accept SYN packets from the other end. This means that the

Chapter 5 67
edge in the state diagram from SYN SENT to SYN RECD should be re-
moved. Instead, the response to SYN in SYN SENT would be something
like RST.
(b) As long as either side is allowed to close, no. The timings of the close()
calls are an application-level issue. If both sides happen to request at ap-
proximately the same instant that a connection be closed, it hardly seems
appropriate to hold the connection open while the requests are serialized.
Looked at another way, disallowing simultaneous opens in effect simply re-
quires that both sides adhere to the established roles of “client” and “server”.
At the point of closing, however, there simply is no established role for
whether client or server is to close the connection. It would indeed be pos-
sible to require that the client, for example, had to initiate the close, but that
would leave the server somewhat at the mercy of the client.
(c) The minimum additional header information for this new interpretation is
a bit indicating whether the sender of the packet was the client for that
connection or the server. This would allow the receiving host to figure out
to which connection the arriving packet belongs.
With this bit in place, we can label the nodes and edges at or above the
ESTABLISHED state with “client” or “server” roles. The edge from LIS-
TEN to SYN SENT is the exception, traversed only if a server (LISTEN
state) takes on a client role (SYN SENT). We replace this edge with a no-
tion of creating the second connection; the original endpoint remains in the
server-role LISTEN state and a new endpoint (with same port number), on
in effect a new diagram, is created in the client-role SYN SENT state.
The edge from SYN SENT to SYN RCVD would be eliminated; during a
simultaneous open the arriving SYN would be delivered to the server-role
endpoint, still in the LISTEN state, rather than to the client-role endpoint
in state SYN SENT.
36. (a) One would now need some sort of connection number assigned by the client
side to play the role of the port number in demultiplexing traffic; with this in
place, headers might not change much at all. Client and server sockets will
now be fundamentally different objects. Server sockets would be required
to bind() themselves to a port number (perhaps at creation time); clients
would be forbidden to do this.
(b) We still need to make sure that a client connection number is not reused
within the 2×MSL period, at least not with the same server port. However,
this is now a TCP-layer responsibility, not an application concern. Assum-
ing the client connection number were assigned at the time of connection,
clients would not need to be aware of TIME WAIT at all: they would be
freed of the requirement they close one socket and reopen a new one to get
a fresh port number.
Since client connection numbers are now not visible to the client, simply
placing a connection number out of service entirely during the TIME WAIT
interval, for connections to any server, would be a tolerable approach.

Chapter 5 68
(c) The rlogin/rsh protocol authenticates clients by seeing that they are using
a “reserved” port on the sending host (normally, a port only available to
system-level processes). This would no longer be possible.
However, the following variation would still be possible: when an rsh
server host S receives a client request from host C, with connection num-
ber N, then S could authenticate the request with C by initiating a second
connection to a reserved port on C, whereupon some sort of authentication
application on C would verify that connection number N was indeed being
used by an authorized rsh client on C. Note that this scheme implies that
connection numbers are at least visible to the applications involved.
37. (a) A program that connect()s, and then sends whatever is necessary to get the
server to close its end of the connection (eg the string “QUIT”), and then
sits there, idle but not disconnecting, will suffice. Note that the server has
to be willing to initiate the active close based on some client action.
(b) Alas, most telnet clients do not work here. Although many can connect to
an arbitrary port, and issue a command such as QUIT to make the server
initiate the close, they generally do close immediately in response to re-
ceiving the server’s FIN.
However, the sock program, written by W. Richard Stevens, can be used
instead. In the (default) client mode, it behaves like a command-line telnet.
The option -Q 100 makes sock wait 100 seconds after receiving the server
FIN before it closes its end of the connection. Thus the command
sock -Q 100 hostname 25
can be used to demonstrate FIN WAIT 2 with an SMTP (email) server
(port 25) on hostname, using the QUIT command.
sock is available from http://www.icir.org/christian/sock.html
38. Let A be the closing host and B the other endpoint. A sends message1, pauses,
sends message2, and then closes its end of the connection for reading. B gets
message1 and sends a reply, which arrives after A has performed the half-close.
B doesn’t read message2 immediately; it remains in the TCP layer’s buffers. B’s
reply arrives at A after the latter has half-closed, and so A responds with RST
as per the quoted passage from RFC 1122. This RST then arrives at B, which
aborts the connection and the remaining buffer contents (i.e. message2) are lost.
Note that if A had performed a full-duplex close, the same scenario can occur.
However, it now depends on B’s reply crossing A’s FIN in the network. The
half-close-for-reading referred to in this exercise is actually purely a local state
change; a connection that performs a half-close closing its end for writing may
however send a FIN segment to indicate this state to the other endpoint.
39. Incrementing the Ack number for a FIN is essential, so that the sender of the FIN
can determine that the FIN was received and not just the preceding data.
For a SYN, any ACK of subsequent data would increment the acknowledgment
number, and any such ACK would implicitly acknowledge the SYN as well (data

Chapter 5 69
cannot be ACKed until the connection is established). Thus, the incrementing of
the sequence number here is a matter of convention and consistency rather than
design necessity.
40. (a) One method would be to invent an option to specify that the first n bytes of
the TCP data should be interpreted as options.
(b) A TCP endpoint receiving an unknown option might
• close/abort the connection. This makes sense if the connection cannot
meaningfully continue when the option isn’t understood.
• ignore the option but keep the TCP data. This is the current RFC 1122
requirement.
• send back “I don’t understand”. This is simply an explicit form of the
previous response. A refinement might be to send back some kind of
list of options the host does understand.
• discard the accompanying the TCP data. One possible use might be
if the data segment were encrypted, or in a format specified by the
option. Some understanding would be necessary regarding sequence
numbers for this to make sense; if the entire TCP data segment was an
extended option block then the sequence numbers shouldn’t increase
at all.
• discard the first n bytes of the TCP data. This is an extension of the
previous strategy to handle the case where the first n bytes of the TCP
data was to be interpreted as an expanded options block; it is not clear
though when the receiver might understand n but not the option itself.
41. TCP faces two separate crash-and-reboot scenarios: a crash can occur in the
middle of a connection, or between two consecutive incarnations of a connection.
The first leads to a “half-open” connection where one endpoint has lost all state
regarding the connection; if either the stateless side sends a new SYN or the
stateful side sends new data, the other side will respond with RST and the half-
open connection will be dissolved bilaterally.
If one host crashes and reboots between two consecutive connection incarna-
tions, the only way the first incarnation could affect the second is if a late-arriving
segment from the first happens to fit into the receive window of the second. TCP
establishes a quasi-random initial sequence number during its three-way hand-
shake at connection open time. A 64KB window, the maximum allowed by the
original TCP, spans less than 0.0015% of the sequence number space. Therefore,
there is very little chance that data from a previous incarnation of the connection
will happen to fall in the current window; any data outside the window is dis-
carded. (TCP also is supposed to implement “quiet time on startup”, an initial
1×MSL delay for all connections after bootup.)
42. (a) Non-exclusive open, reading block N, writing block N, and seeking to
block N all are idempotent, i.e. have the same effect whether executed
once or twice.

Chapter 5 70
(b) create() is idempotent if it means “create if nonexistent, or open if it exists
already”. mkdir() is idempotent if the semantics are “create the given di-
rectory if it does not exist; otherwise do nothing”. delete() (for either file
or directory) works this way if its meaning is “delete if the object is there;
otherwise, ignore.”
Operations fundamentally incompatible with at-least-once semantics in-
clude exclusive open (and any other form of file locking), and exclusive
create.
(c) The directory-removing program would first check if the directory exists.
If it does not, it would report its absence. If it does exist, it invokes the
system call rmdir().
43. (a) The problem is that reads aren’t serviced in FIFO order; disk controllers
typically use the “elevator” or SCAN algorithm to schedule writes, in which
the pool of currently outstanding writes is sorted by disk track number and
the writes are then executed in order of increasing track number. Using a
single channel would force writes to be executed serially even when such a
sequence required lots of otherwise-unnecessary disk head motion.
If a pool of N sequential channels were used, the disk controller would at
any time have about N writes to schedule in the order it saw fit.
(b) Suppose a client process writes some data to the server, and then the client
system shuts down “gracefully”, flushing its buffers (or avails itself of some
other mechanism to flush the buffer cache). At this point data on a local
disk would be safe; however, a server crash would now cause the loss of
client data remaining in the server’s buffers. The client might never be able
to verify that the data was safely written out.
(c) One approach would be to modify a protocol that uses sequential channels
to support multiple independent outstanding requests on a single logical
channel, and to support replies in an arbitrary order, not necessarily that in
which the corresponding requests were received. Such a mechanism would
allow the server to respond to multiple I/O requests in whatever order was
most convenient.
A subsequent request could now no longer serve as an ACK of a previous
reply; ACKs would have to be explicit and non-cumulative. There would
be changes in retransmission management as well: the client would have
to maintain a list of the requests that hadn’t yet been answered and the
server would have to maintain a list of replies that had been sent but not
acknowledged. Some bound on the size of these lists (corresponding to
window size) would be necessary.
44. (a) The client sends the request. The server executes it (and successfully com-
mits any resulting changes to disk), but then crashes just before sending its
reply. The client times out and resends the request, which is executed a
second time by the server as it restarts.

Chapter 5 71
(b) The tipoff to the client that this might have happened is that the server’s
boot ID field incremented over that from the previous request (which would
always cause the RPC call to fail). While a server reboot would always be
indicated by an incremented boot ID, it would not necessarily be the case
that any particular request was actually executed twice.
45. We will use the log blocks to maintain a “transaction log”, a simplified version of
the strategy used by database applications. In this particular example the actual
update is atomic; if two data blocks had to be updated together we would have
additional complications.
Upon receipt of the request, the RPC server does the following:
• reads in block N from the disk.
• records in the log block the CID and MID values, the values of X and N,
and an indication that the transaction is in progress.
• performs the actual update write.
• replaces the log entry with one that contains CID and MID and an indication
that the operation was successful, and sends the reply stating this.
This last logfile record is retained until the client ACKs the reply.
On restart the server looks in the log block. If this indicates nothing about the
transaction, then either the transaction was never started or else the final ACK
was received; either way, the RPC server has no further immediate responsibil-
ities. If the log block indicates that the transaction completed successfully, we
reload its status as completed but unacknowledged. The server doesn’t know
if the reply has been sent, but this doesn’t matter as it will be retransmitted if
necessary when the appropriate timeout occurs. If such a retransmission was
unnecessary, then the client will infer this from the expired MID.
Finally, if the restarting server finds the in-progress indication in the log, then it
reads data block N and determines, by comparing X there with the X in the log,
whether the write operation completed. If so, the log is updated as in the fourth
step above; if not, the server resumes the sequence above at the third step, the
point of performing the actual write.
46. (a) If a client has only sent the request once, and has received a reply, and if the
underlying network never duplicates packets, then the client can be sure its
request was only executed once.
(b) To ensure at-most-once semantics a server would have to buffer a reply with
a given transaction XID until it had received an acknowledgment from the
client that the reply had been received properly. This would entail adding
such ACKs to the protocol, and also adding the appropriate buffering mech-
anism to the implementation.
47. One TCP connection can manage multiple outstanding requests, and so is capa-
ble of supporting multiple logical channels; we will assume that this is the case.
The alternative, of one TCP connection per channel, is similar.

Chapter 5 72
(a) The overlying RPC protocol would need to provide a demultiplexing field
corresponding to the channel ID. (In the one-TCP-connection-per-channel
setting, the TCP socketpair defining the connection represents the channel
ID.)
(b) The message ID would correspond to the sequence number; the primary
purpose of the message ID is to keep track of acknowledgments.
(c) Boot ID is dealt with by the stateful nature of TCP; if either end rebooted
and the other end eventually sent a message, the RST response would be
an indication of that reboot.
(d)(e) The RPC request and reply messages would now become RPC headers
that divide the TCP byte stream into discrete messages. There would be
no guarantee, of course, that these headers were transmitted in the same
segment as the associated data.
(f) The RPC ACK would be replaced by the TCP ACK.
(g) Some sort of are-you-alive? messages would still have to be generated
by the client, if they were desired; although TCP does support KeepAlive
messages they are for a vastly different (∼2-hour) time scale and they do
not address the issue of whether the server process is alive.
If the TCP layer delayed sending ACKs for, say, 100ms (such “Delayed ACKs”
are standard practice), then in many instances the reply might be available in
time for it to be sent with the TCP ACK. This would achieve the effect of implicit
ACKs in having only one packet handle both ACK and reply.
49. • An application that encodes audio or video might produce a group of pack-
ets at a certain time that needed to be spread out in time for appropriate
playback. The application would typically do better sending those pack-
ets when they are ready rather than trying to pace them smoothly into the
network (which could increase total delay).
• An application might send video and audio data at slightly different times
that needed to be synchronized, or a single video frame might be sent in
multiple pieces over time.
It follows from the above that only the application (not the RTP stack or the
network) has the appropriate knowledge of when a particular item should be
played back, and thus the application should provide the timestamps.
50. This allows the server to make accurate measurements of jitter. This in turn
allows an early warning of transient congestion; see the solution to Exercise 53
below. Jitter data might also allow finer control over the size of the playback
buffer, although it seems unlikely that great accuracy is needed here.
51. Each receiver gets 1/1000 of 5% of 320 kbps, or 16bps, which means one 84-
byte RTCP packet every 42 sec. At 10K recipients, it’s one packet per 420 sec,
or 7 minutes.

Chapter 5 73
52. (a) The answer here depends on how closely frame transmission is synchro-
nized with frame display. Assuming playback buffers on the order of a
full frame or larger, it seems likely that receiver frame-display finish times
would not be synchronized with frame transmission times, and thus would
not be particularly synchronized from receiver to receiver. In this case,
receiver synchronization of RTCP reports with the end of frame display
would not result in much overall synchronization of RTCP traffic.
In order to achieve such synchronization, it would be necessary to have both
a very uniform latency for all receivers and a rather low level of jitter, so
that receivers were comfortable maintaining a negligible playback buffer.
It would also be necessary, of course, to disable the RTCP randomization
factor. The number of receivers, however, should not matter.
(b) The probability that any one receiver sends in the designated 5% subinter-
val is 0.05, assuming uniform distribution; the probability that all 10 send
in the subinterval is 0.0510
, which is negligible.
(c) The probability that one designated set of five receivers sends in the desig-
nated interval, and the other five do not, is (.05)5
× (.95)5
. There are (10
choose 5) = 10!/5!5! ways of selecting five designated receivers, and so the
probability that some set of five receivers all transmit in the designated in-
terval is (10 choose 5) ×(.05)5
×(.95)5
= 252×0.0000002418 = 0.006%.
Multiplying by 20 gives a rough estimate of about 0.12% for the probabil-
ity of an upstream traffic burst rivaling the downstream burst, in any given
reply interval.
53. If most receivers are reporting high loss rates, a server might consider throttling
back. If only a few receivers report such losses, the server might offer referrals
to lower-bandwidth/lower-resolution servers. A regional group of receivers re-
porting high losses might point to some local congestion; as RTP traffic is often
tunneled, it might be feasible to address this by re-routing traffic.
As for jitter measurements, we quote RFC 1889:
The interarrival jitter field provides a second short-term measure of
network congestion. Packet loss tracks persistent congestion while
the jitter measure tracks transient congestion. The jitter measure may
indicate congestion before it leads to packet loss.
54. Many answers are possible here. RTT estimation, and hence calculation of suit-
able timeout values, is more difficult than TCP because of the lack of a closed
feedback loop between sender and receiver. The solution could include looking
for gaps in the RTP sequence number space. Running another protocol on top
of RTP (see DCCP, RFC 4340, for example) to detect losses via an acknowledg-
ment mechanism is another option.

1. (a) From the application’s perspective, it is better to define flows as process-
to-process. If a flow is host-to-host, then an application running on a multi-
user machine may be penalized (by having its packets dropped) if another
application is heavily using the same flow. However, it is much easier
to keep track of host-to-host flows; routers need only look at the IP ad-
dresses to identify the flow. If flows are process-to-process (i.e. end-to-
end), routers must also extract the TCP or UDP ports that identify the end-
points. In effect, routers have to do the same demultiplexing that is done
on the receiver to match messages with their flows.
(b) If flows are defined on a host-to-host basis, then FlowLabel would be a
hash of the host-specific information; that is, the IP addresses. If flows are
process-to-process, then the port numbers should be included in the hash
input.
2. (a) In a rate-based TCP the receiver would advertise a rate at which it could
receive data; the sender would then limit itself to this rate, perhaps making
use of a token bucket filter with small bucket depth. Congestion-control
mechanisms would also be converted to terms of throttling back the rate
rather than the window size. Note that a window-based model sending one
window-full per RTT automatically adjusts its rate inversely proportional
to the RTT; a rate-based model might not. Note also that if an ACK arrives
for a large amount of data, a window-based mechanism may immediately
send a burst of a corresponding large amount of new data; a rate-based
mechanism would likely smooth this out.
(b) A router-centric TCP would send as before, but would receive (presumably
a steady stream of) feedback packets from the routers. All routers would
have to participate, perhaps through a connection-oriented packet-delivery
model. TCP’s mechanisms for inferring congestion from changes in RTT
would all go away.
TCP might still receive some feedback from the receiver about its rate, but
the receiver would only do so as a “router” of data to an application; this is
where flow control would take place.
3. For Ethernet, throughput with N stations is 5/(N/2 + 5) = 10/(N + 10); to
send one useful packet we require N/2 slots to acquire the channel and 5 slots
to transmit. On average, a waiting station has to wait for about half the others
to transmit first, so with N stations the delay is the time it takes for N/2 to
transmit; combining this with a transmission time of N/2 + 5 this gives a total
delay of N/2 × (N/2 + 5) = N(N + 10)/4. Finally, power is throughput/delay
= 40/N(N + 10)2
. Graphs are below.
75

Chapter 6 76
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10
power
throughput
0
10
20
30
40
50
delay
delay
throughput,
power
Load, N
Ethernet throughput, delay, power (Exercise 3(a))
The power curves has its maximum at N = 1, the minimum N, which is some-
what artificial and is an artifact of the unnatural way we are measuring load.
4. Throughput here is min(x, 1), where x is the load. For x ≤ 1 the delay is 1
second, constantly. We cannot sustain x 1 at all; the delay approaches infinity.
The power curve thus looks like y = x for x ≤ 1 and is undefined beyond that.
Another way to measure load might be in terms of the percentage of time the
peak rate exceeds 1, assuming that the average rate remains less than 1.
5. Yes, particularly if the immediate first link is high-bandwidth, the first router has
a large buffer capacity, and the delay in the connection is downstream. Conges-
tionWindow can grow arbitrarily; the excess packets will simply pile up at the
first router.
6. R1 cannot become congested because traffic arriving at one side is all sent out
the other, and the bandwidths on each side are the same.
We now show how to congest only the router R2 that is R1’s immediate left child;
other R’s are similar.

Chapter 6 77
R1
R
R2
R R3 R R
H H H1 H2 H3 H4 H5 H6
Link bandwidth 4 MB/sec
We arrange for H3 and H4 to send 1MB/sec to H1, and H5 and H6 to send
1MB/sec to H2. Each of the links to the right of R1 reaches its maximum ca-
pacity, as does the R1—R2 link, but none of these routers becomes congested.
However, R2 now wants to send 4MB/sec to R3, which it cannot.
R3 is not congested as it receives at 2MB/sec from R2 and this traffic is evenly
divided between H1 and H2.
7. (a) The fairness index is 0.9360; x1 + · · · + x5 = 715 and x2
1 + · · · + x2
5 =
109225.
(b) The index falls to 0.4419.
8. Fi still represents a timestamp, but now when computing Fi as a packet arrives
we run the clock slow by the sum of the weights w of the active flows, rather
than by the number of active flows.
Consider two flows with weights 1 and 2. If the the packet size of the packet
in the queue for flow 2 is twice that of the packet in flow 1, then both packets
should look equally attractive to transmit. Hence, the effective packet size of the
second packet should be Pi/2. In general, if the flow has a weight w then the
effective packet size is Pi/w. Hence the final time-stamps are calculated as
Fi = max(Fi−1, Ai) + Pi/w
9. If we are in the process of transmitting a large sized packet and a small packet
arrives just after the start of the transmission, then due to non-preemption the
small packet gets transmitted after the large. However, in perfect bit-by-bit round
robin the small packet would have finished being transmitted before the large
packet gets completely transmitted.
10. (a) First we calculate the finishing times Fi. We don’t need to worry about
clock speed here since we may take Ai = 0 for all the packets. Fi thus
becomes just the cumulative per-flow size, i.e. Fi = Fi−1 + Pi.

Chapter 6 78
Packet size flow Fi
1 100 1 100
2 100 1 200
3 100 1 300
4 100 1 400
5 190 2 190
6 200 2 390
7 110 3 110
8 50 3 170
We now send in increasing order of Fi:
Packet 1, Packet 7, Packet 8, Packet 5, Packet 2, Packet 3, Packet 6, Packet
4.
(b) To give flow 2 a weight of 4 we divide each of its Fi by 4, i.e. Fi =
Fi−1 + Pi/4; again we are using the fact that there is no waiting.
Packet size flow weighted Fi
1 100 1 100
2 100 1 200
3 100 1 300
4 100 1 400
5 190 2 47.5
6 200 2 97.5
7 110 3 110
8 50 3 170
Transmitting in increasing order of the weighted Fi we send as follows:
Packet 5, Packet 6, Packet 1, Packet 7, Packet 8, Packet 2, Packet 3, Packet
4.
12. (a) The advantage would be that the dropped packets are the resource hogs, in
terms of buffer space consumed over time. One drawback is the need to
recompute cost whenever the queue advances.
(b) Suppose the queue contains three packets. The first has size 5, the second
has size 15, and the third has size 5. Using the sum of the sizes of the earlier
packets as the measure of time remaining, the cost of the third packet is
5 × 20 = 100, and the cost of the (larger) second is 15 × 5 = 75. (We
have avoided the issue here of whether the first packet should always have
cost 0, which might be mathematically correct but is arguably a misleading
interpretation.)
(c) We again measure cost in terms of size; i.e. we assume it takes 1 time unit
to transmit 1 size unit. A packet of size 3 arrives at T=0, with the queue
such that the packet will be sent at T=5. A packet of size 1 arrives right
after.
At T=0 the costs are 3 × 5 = 15 and 1 × 8 = 8.
At T=3 the costs are 3 × 2 = 6 and 1 × 5 = 5.

Chapter 6 79
At T=4 the costs are 3×1 = 3 and 1×4 = 4; cost ranks have now reversed.
At T=5 the costs are 0 and 3.
13. (a) With round-robin service, we will alternate one telnet packet with each ftp
packet, causing telnet to have dismal throughput.
(b) With FQ, we send roughly equal volumes of data for each flow. There are
about 552/41 ≈ 13.5 telnet packets per ftp packet, so we now send 13.5
telnet packets per ftp packet. This is better.
(c) We now send 512 telnet packets per ftp packet. This excessively penalizes
ftp.
Note that with the standard Nagle algorithm a backed-up telnet would not in fact
send each character in its own packet.
14. In light of the complexity of the solution here, instructors may wish to consider
limiting the exercise to those packets arriving before, say, T=6.
(a) For the ith arriving packet on a given flow we calculate its estimated finish-
ing time Fi by the formula Fi = max{Ai, Fi−1}+1, where the clock used
to measure the arrival times Ai runs slow by a factor equal to the number
of active queues. The Ai clock is global; the sequence of Fi’s calculated
as above is local to each flow. A helpful observation here is that packets
arrive and are sent at integral wallclock times.
The following table lists all events by wallclock time. We identify packets
by their flow and arrival time; thus, packet A4 is the packet that arrives on
flow A at wallclock time 4, i.e. the third packet. The last three columns
are the queues for each flow for the subsequent time interval, including
the packet currently being transmitted. The number of such active queues
determines the amount by which Ai is incremented on the subsequent line.
Multiple packets appear on the same line if their Fi values are all the same;
the Fi values are in italic when Fi = Fi−1 + 1 (versus Fi = Ai + 1).
We decide ties in the order flow A, flow B, flow C. In fact, the only ties are
between flows A and C; furthermore, every time we transmit an A packet
we have a C packet tied with the same Fi.

Chapter 6 80
Wallclock Ai arrivals Fi sent A’s queue B’s queue C’s queue
1 1.0 A1,C1 2.0 A1 A1 C1
2 1.5 B2 2.5 C1 A2 B2 C1,C2
A2,C2 3.0
3 1.833 C3 4.0 B2 A2 B2 C2,C3
4 2.166 A4 4.0 A2 A2,A4 C2,C3
5 2.666 C5 5.0 C2 A4 C2,C3,C5
6 3.166 A6 5.0 A4 A4,A6 B6 C3,C5,C6
B6 4.166
C6 6.0
7 3.5 A7 6.0 C3 A6,A7 B6 C3,C5,C6,C7
C7 7.0
8 3.833 B8 5.166 B6 A6,A7 B6,B8 C5,C6,C7,C8
C8 8.0
9 4.166 A9 7.0 A6 A6,A7,A9 B8 C5,C6,C7,C8
10 4.5 A10 8.0 C5 A7,A9,A10 B8 C5,C6,C7,C8
11 4.833 B11 6.166 B8 A7,A9,A10 B8,B11 C6,C7,C8
12 5.166 B12 7.166 A7 A7,A9,A10 B11,B12 C6,C7,C8
13 5.5 C6 A9,A10 B11,B12 C6,C7,C8
14 5.833 B11 A9,A10 B11,B12 C7,C8
15 6.166 B15 8.166 A9 A9,A10 B12,B15 C7,C8
16 C7 A10 B12,B15 C7,C8
17 B12 A10 B12,B15 C8
18 A10 A10 B15 C8
19 C8 B15 C8
20 B15 B15
(b) For weighted fair queuing we have, for flow C,
Fi = max{Ai, Fi−1} + 0.5
For flows A and B, Fi is as before. Here is the table corresponding to the
one above.

Chapter 6 81
Wallclock Ai arrivals Fi sent A’s queue B’s queue C’s queue
1 1.0 A1 2.0 C1 A1 C1
C1 1.5
2 1.5 A2 3.0 A1 A1,A2 B2 C2
B2 2.5
C2 2.0
3 1.833 C3 2.5 C2 A2 B2 C2,C3
4 2.166 A4 4.0 B2 A2,A4 B2 C3
5 2.5 C5 3.0 C3 A2,A4 C3,C5
6 3.0 A6 5.0 A2 A2,A4,A6 B6 C5,C6
B6 4.0
C6 3.5
7 3.333 A7 6.0 C5 A4,A6,A7 B6 C5,C6,C7
C7 4.0
8 3.666 B8 5.0 C6 A4,A6,A7 B6,B8 C6,C7,C8
C8 4.5
9 4.0 A9 7.0 A4 A4,A6,A7,A9 B6,B8 C7,C8
10 4.333 A10 8.0 B6 A6,A7,A9,A10 B6,B8 C7,C8
11 4.666 B11 6.0 C7 A6,A7,A9,A10 B8,B11 C7,C8
12 5.0 B12 7.0 C8 A6,A7,A9,A10 B8,B11,B12 C8
13 5.333 A6 A6,A7,A9,A10 B8,B11,B12
14 5.833 B8 A7,A9,A10 B8,B11,B12
15 6.333 B15 8.0 A7 A7,A9,A10 B11,B12,B15
16 B11 A9,A10 B11,B12,B15
17 A9 A9,A10 B12,B15
18 B12 A10 B12,B15
19 A10 A10 B15
20 B15 B15
16. (a) In slow start, the size of the window doubles every RTT. At the end of the
ith RTT, the window size is 2i
KB. It will take 10 RTTs before the send
window has reached 210
KB = 1 MB.
(b) After 10 RTTs, 1023 KB = 1 MB − 1 KB has been transferred, and the
window size is now 1 MB. Since we have not yet reached the maximum
capacity of the network, slow start continues to double the window each
RTT, so it takes 4 more RTTs to transfer the remaining 9MB (the amounts
transferred during each of these last 4 RTTs are 1 MB, 2 MB, 4 MB, 1 MB;
these are all well below the maximum capacity of the link in one RTT of
12.5 MB). Therefore, the file is transferred in 14 RTTs.
(c) It takes 0.7 seconds (14 RTTs) to send the file. The effective throughput
is (10MB / 0.7s) = 14.3MBps = 114.3Mbps. This is only 11.4% of the
available link bandwidth.

Chapter 6 82
17. Let the sender window size be 1 packet initially. The sender sends an entire
window-full in one batch; for every ACK of such a window-full that the sender
receives, it increases its effective window (which is counted in packets) by one.
When there is a timeout, the effective window is cut into half the number of
packets.
Now consider the situation when the indicated packets are lost. The window size
is initially 1; when we get the first ACK it increases to 2. At the beginning of
the second RTT we send packets 2 and 3. When we get their ACKs we increase
the window size to 3 and send packets 4, 5 and 6. When these ACKs arrive the
window size becomes 4.
Now, at the beginning of the fourth RTT, we send packets 7, 8, 9, and 10; by
hypothesis packet 9 is lost. So, at the end of the fourth RTT we have a timeout
and the window size is reduced to 4/2 = 2.
Continuing, we have
RTT 5 6 7 8 9
Sent 9-10 11-13 14-17 18-22 23-28
Again the congestion window increases up until packet 25 is lost, when it is
halved, to 3, at the end of the ninth RTT. The plot below shows the window size
vs. RTT.
CONGESTION WINDOW
SIZE
RTT
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
1
2
3
4
5
6
7
18. From the figure for the preceding exercise we see that it takes about 17 RTTs
for 50 packets, including the necessary retransmissions. Hence the effective
throughput is 50/17 × 100 × 10−3
KB/s = 29.4 KB/s.
19. The formula is accurate if each new ACK acknowledges one new MSS-sized
segment. However, an ACK can acknowledge either small size packets (smaller
than MSS) or cumulatively acknowledge many MSS’s worth of data.

Chapter 6 83
Let N = CongestionWindow/MSS, the window size measured in segments.
The goal of the original formula was so that after N segments arrived the net
increment would be MSS, making the increment for one MSS-sized segment
MSS/N. If instead we receive an ACK acknowledging an arbitrary Amoun-
tACKed, we should thus expand the window by
Increment = AmountACKed/N
= (AmountACKed × MSS)/CongestionWindow
20. We may still lose a batch of packets, or else the window size is small enough that
three subsequent packets aren’t sent before the timeout. Fast retransmit needs
to receive three duplicate ACKs before it will retransmit a packet. If so many
packets are lost (or the window size is so small) that not even three duplicate
ACKs make it back to the sender, then the mechanism cannot be activated, and a
timeout will occur.
21. We will assume in this exercise and the following two that when TCP encoun-
ters a timeout it reverts to stop-and-wait as the outstanding lost packets in the
existing window get retransmitted one at a time, and that the slow start phase
begins only when the existing window is fully acknowledged. In particular, once
one timeout and retransmission is pending, subsequent timeouts of later packets
are suppressed or ignored until the earlier acknowledgment is received. Such
timeouts are still shown in the tables below, but no action is taken.
We will let Data N denote the Nth packet; Ack N here denotes the acknowledg-
ment for data up through and including data N.
(a) Here is the table of events with TimeOut = 2 sec. There is no idle time on
the R–B link.
Time A recvs A sends R sends cwnd size
0 Data0 Data0 1
1 Ack0 Data1,2 Data1 2
2 Ack1 Data3,4 (4 dropped) Data2 3
4 Ack3/timeout4 Data 4 Data5 1
5 Ack3/timeout56 Data4 1
6 Ack5 Data 6 Data6 1
7 Ack 6 Data7,8 (slow start) Data7 2
(b) With TimeOut = 3 sec, we have the following. Again nothing is transmitted
at T=6 because ack 4 has not yet been received.

Chapter 6 84
Time A recvs A sends R sends cwnd size
0 Data0 Data0 1
1 Ack0 Data1,2 Data1 2
5 Ack3/timeout4 Data4 Data7 1
6 Ack3/timeout56 Data4 1
7 Ack5/timeout78 Data6 Data6 1
8 Ack7 Data8 Data8 1
9 Ack8 Data9,10 (slow start) Data9 2
22. We follow the conventions and notation of the preceding exercise. Although the
first packet is lost at T=4, it wouldn’t have been transmitted until T=8 and its loss
isn’t detected until T=10. During the final few seconds the outstanding losses in
the existing window are made up, at which point slow start would be invoked.
A recvs cwnd A sends R sending/R’s queue
Ack # size Data
T=0 1 1 1/
T=1 1 2 2,3 2/3
T=2 2 3 4,5 3/4,5
T=3 3 4 6,7 4/5,6,7
T=4 4 5 8,9 5/6,7,8 9 lost
T=5 5 6 10,11 6/7,8,10 11 lost
T=6 6 7 12,13 7/8,10,12 13 lost
T=7 7 8 14,15 8/10,12,14 15 lost
T=8 8 9 16,17 10/12,14,16 17 lost
T=9 8 9 12/14,16
T=10 8 9 9 14/16,9 2nd duplicate Ack8
T=11 8 16/9
T=12 8 9/
T=13 10 11 11/ B gets 9
T=14 12 13 13/
T=15 14 15 15/
T=16 16 17 17/
T=17 17 2 18,19 18/19 slow start
23. R’s queue size is irrelevant because the R-B link changed from having a band-
width delay to having a propagation delay only. That implies that packets leave
R as soon as they arrive and hence no queue can develop. The problem now be-
comes rather trivial compared to the two previous questions. Because no queue
can develop at the router, packets will not be dropped, so the window continues
to grow each RTT. In reality this scenario could happen but would ultimately be
limited by the advertised window of the connection.
Note that the question is somewhat confusingly worded—it says that 2 packets

Chapter 6 85
take one second to send, but since this is propagation delay rather than bandwidth
delay, any number of packets can be sent in one second.
Notation and conventions are again as in #21 above.
A recvs cwnd A sends data #
T=0 1 1
T=1 Ack1 2 2,3
T=2 Ack3 4 4,5,6,7
T=3 Ack7 8 8–15
T=4 Ack15 16 16–31
T=5 Ack31 32 32–63
T=6 Ack63 64 64–127
T=7 Ack127 128 127–255
T=8 Ack255 256 255–511
24. With a full queue of size N, it takes an idle period on the sender’s part of N+1
seconds for R1’s queue to empty and link idling to occur. If the connection is
maintained for any length of time with CongestionWindow=N, no losses occur
but EstimatedRTT converges to N. At this point, if a packet is lost the timeout
of 2×N means an idle stretch of length 2N − (N+1) = N−1.
With fast retransmit, this idling would not occur.
25. The router is able in principle to determine the actual number of bytes outstand-
ing in the connection at any time, by examining sequence and acknowledgment
numbers. This we can take to be the congestion window except for immediately
after when the latter decreases.
The host is complying with slow start at startup if only one more packet is out-
standing than the number of ACKs received. This is straightforward to measure.
Slow start after a coarse-grained timeout is trickier. The main problem is that the
router has no way to know when such a timeout occurs; the TCP might have in-
ferred a lost packet by some other means. We may, however, on occasion be able
to rule out three duplicate ACKs, or even two, which means that a retransmission
might be inferred to represent a timeout.
After any packet is retransmitted, however, we should see the congestion window
fall at least in half. This amounts to verifying multiplicative decrease, though,
not slow start.
26. Using ACKs in this manner allow very rapid increase and control over Conges-
tionWindow. Stefan Savage suggests requiring ACKS to include a nonce as a
solution. That is, ACKs must include information from that data which is being
ACKed to be valid.
27. Slow start is active up to about 0.5 sec on startup. At that time a packet is sent
that is lost; this loss results in a coarse-grained timeout at T=1.9.
At that point slow start is again invoked, but this time TCP changes to the linear-
increase phase of congestion avoidance before the congestion window gets large

Chapter 6 86
enough to trigger losses. The exact transition time is difficult to see in the dia-
gram; it occurs sometime around T=2.4.
At T=5.3 another packet is sent that is lost. This time the loss is detected at T=5.5
by fast retransmit; this TCP feature is the one not present in Figure 6.11 of the
text, as all lost packets there result in timeouts. Because the congestion window
size then drops to 1, we can infer that fast recovery was not in effect; instead,
slow start opens the congestion window to half its previous value and then linear
increase takes over. The transition between these two phases is shown more
sharply here, at T=5.7.
28. We assume here that the phone link delay is due to bandwidth, not latency, and
that the rest of the network path offers a bandwidth at least as high as the phone
link’s. During the first RTT we send one packet, due to slow start, and by the final
assumption we thus transmit over the link for a third of the RTT, and thus use
only a third of the total bandwidth, or 1 KB/sec. During the second RTT we send
two packets; in the third and subsequent RTTs send three packets, saturating the
phone link. The sequence of averages, however, climbs more slowly: at the end
of the second RTT the fraction of bandwidth used is 3/6; at the end of the third
RTT it is 6/9, then 9/12, at the end of the Nth RTT we have used 1-1/N of the
bandwidth.
Packet losses cause these averages to drop now and then, although since the
averages are cumulative the drops are smaller and smaller as time goes on.
29. (a) Here is how a connection startup might progress:
Send packet 1
Get ack 1
Send packets 2 3
Get ack 2
Send packet 4, which is lost due to link errors, so CongestionWin-
dow=1.
One way or another, we get lots of coarse-grained timeouts when the win-
dow is still too small for fast retransmit. We will never be able to get past
the early stages of slow start.
(b) Over the short term such link losses cannot be distinguished from conges-
tion losses, unless some router feedback mechanism (e.g. ICMP Source
Quench) were expanded and made more robust. (Over the long term, con-
gestion might be expected to exhibit greater temporal variability, and care-
ful statistical analysis might indicate when congestion was present.)
(c) In the presence of explicit congestion indications, TCP might now be tuned
to respond to ordinary timeout losses by simply retransmitting, without
reducing the window size. Large windows could now behave normally.
We would, however, need some way for keeping the ACK clocking run-
ning; coarse-grained timeouts would still necessitate a return to
CongestionWindow= 1 because ACKs would have drained. Either TCP’s
existing fast retransmit/fast recovery, or else some form of selective ACKs,

Chapter 6 87
might be appropriate. Either might need considerable tuning to handle a
25% loss rate.
30. Suppose the first two connections keep the queue full 95% of the time, alternating
transmissions in lockstep and timed so that their packets always arrive just as a
queue vacancy opens. Suppose also that the third connection’s packets happen
always to arrive when the queue is full. The third connection’s packets will thus
be lost, whether we use slow start or not. The first two connections will not be
affected.
Congestion avoidance by the first two connections means that they will even-
tually try a window size of 4, and fall back to 2, and give the third connection
a real foot in the door. Slow start for the third connection would mean that if
a packet got through, then the window would expand to 2 and the third sender
would have about twice the probability of getting at least one packet through.
However, since a loss is likely, the window size would soon revert to 1.
31. (a) We lose 1100 ms: we wait 300 ms initially to detect the third duplicate
ACK, and then one full 800 ms RTT as the sender waits for the ACK of the
retransmitted segment. If the lost packet is sent at T=−800, the lost ACK
would have arrived at T=0. The duplicates arrive at T=100, 200, and 300.
We retransmit at T=300, and the ACK finally arrives at T=1100.
(b) We lose 1100 − 400 = 700 ms. As shown in the diagram, the elapsed time
before we resume is again 1100 ms but we have had four extra chances to
transmit during that interval, for a savings of 400 ms.

Chapter 6 88
-100
0
100
200
300
400
500
600
700
800
900
1000
1100
3rd dup ACK
data
ack
retransmitted data
resume
stop sending
-200
-300
-400
-500
-600
-700
-800
-100
0
100
200
300
400
500
600
700
800
900
1000
1100
3rd dup ACK
data
ack
retransmitted data
resume
stop sending
-200
-300
-400
-500
-600
-700
-800
29(a) 29(b)
window forward
window forward
window forward
window forward
32. We might alternate between congestion-free backoff and heavy congestion, mov-
ing from the former to the latter in as little as 1 RTT. Moving from congestion
back to no congestion unfortunately tends not to be so rapid.
TCP Reno also oscillates between congestion and non-congestion, but the peri-
ods of non-congestion are considerably longer.
33. Marking a packet allows the endpoints to adjust to congestion more efficiently—
they may be able to avoid losses (and timeouts) altogether by slowing their send-
ing rates. However, transport protocols must be modified to understand and ac-
count for the congestion bit. Dropping packets leads to timeouts, and therefore
may be less efficient, but current protocols (such as TCP) need not be modified
to use RED. Also, dropping is a way to rein in an ill-behaved sender.
34. (a) We have
TempP = MaxP ×
AvgLen−MinThreshold
MaxThreshold−MinThreshold
.
AvgLen is halfway between MinThreshold and MaxThreshold, which
implies that the fraction here is 1/2 and so TempP = MaxP/2 = 0.005.
We now have Pcount = TempP/(1 − count × TempP) = 1/(200−count).
For count=1 this is 1/199; for count=100 it is 1/100.

Chapter 6 89
(b) Evaluating the product (1 − P1) × · · · × (1 − P50) gives
198
199
×
197
198
×
196
197
× · · · ×
150
151
×
149
150
which all telescopes down to 149/199, or 0.7487.
36. The difference between MaxThreshold and MinThreshold should be large enough
to accommodate the average increase in the queue length in one RTT; with TCP
we expect the queue length to double in one RTT, at least during slow start, and
hence want MaxThreshold to be at least twice MinThreshold. MinThresh-
old should also be set at a high enough value so that we extract maximum link
utilization. If MaxThreshold is too large, however, we lose the advantages of
maintaining a small queue size; excess packets will simply spend time waiting.
37. Only when the average queue length exceeds MaxThreshold are packets au-
tomatically dropped. If the average queue length is less than MaxThreshold,
incoming packets may be queued even if the real queue length becomes larger
than MaxThreshold. The router must be able to handle this possibility.
38. It is easier to allocate resources for an application that can precisely state its
needs, than for an application whose needs vary over some range. Bursts con-
sume resources, and are hard to plan for.
39. Between MinThreshold and MaxThreshold we are using the drop probability
as a signaling mechanism; a small value here is sufficient for the purpose and a
larger value simply leads to multiple packets dropped per TCP window, which
tends to lead to unnecessarily small window sizes.
Above MaxThreshold we are no longer signaling the sender. There is no logical
continuity intended between these phases.
40. The bit allows for incremental deployment, in which some endpoints respond to
congestion marks and some do not. Without this bit, ECN-enabled routers would
mark packets during congestion rather than dropping them, but some (presum-
ably older, not updated) endpoints would not recognize the mark, and hence
would not back off during congestion, crowding out the ECN-compliant end-
points, which would then have the incentive to ignore ECN marks as well. The
result could actually be congestion collapse as in the pre-congestion-controlled
Internet.
41. (a) Assume the TCP connection has run long enough for a full window to be
outstanding (which may never happen if the first link is the slowest). We
first note that each data packet triggers the sending of exactly one ACK,
and each ACK (because the window size is constant) triggers the sending
of exactly one data packet.
We will show that two consecutive RTT-sized intervals contain the same
number of transmissions. Consider one designated packet, P1, and let the

Chapter 6 90
first RTT interval be from just before P1 is sent to just before P1’s ACK,
A1, arrives. Let P2 be the data packet triggered by the arrival of A1, let
A2 be the ACK for P2, and let the second interval be from just before the
sending of P2 to just before the receipt of A2. Let N be the number of
segments sent within the first interval, i.e., counting P1 but not P2. Then,
because packets don’t cross, this is the number of ACKs received during
the second RTT interval, and these ACKs trigger the sending of exactly N
segments during the second interval as well.
(b) The following shows a window size of four, but only two packets sent per
RTT once the steady state is reached. It is based on an underlying topology
A—R—B, where the A–R link has infinite bandwidth and the R–B link
sends one packet per second each way. We thus have RTT=2 sec; in any
2-second interval beginning on or after T=2 we send only two packets.
T=0 send data[1] through data[4]
T=1 data[1] arrives at destination; ACK[1] starts back
T=2 receive ACK[1], send data[5]
The extra packets are, of course, piling up at the intermediate router.
42. The first time a timed packet takes the doubled RTT, TCP Vegas still sends one
windowful and so measures an ActualRate = CongestionWindow/RTTnew of
half of what it had been, and thus about half (or less) of ExpectedRate. We
then have Diff = ExpectedRate−ActualRate ≈ (1/2)×ExpectedRate, which
is relatively large (and, in particular, larger than β), so TCP Vegas starts reducing
CongestionWindow linearly. This process stops when Diff is much closer to 0;
that is, when CongestionWindow has shrunk by a factor close to two.
The ultimate effect is that we underestimate the usable congestion window by
almost a factor of two.
43. (a) If we send 1 packet, then in either case we see a 1 sec RTT. If we send a
burst of 10 packets, though, then in the first case ACKs are sent back at
1 sec intervals; the last packet has a measured RTT of 10 sec. The second
case gives a 1 sec RTT for the first packet and a 2 sec RTT for the last.
The technique of packet-pairs, sending multiple instances of two consecu-
tive packets right after one another and analyzing the minimum time differ-
ence between their ACKs, achieves the same effect; indeed, packet-pair is
sometimes thought of as a technique to find the minimum path bandwidth.
In the first case, the two ACKs of a pair will always be 1 second apart; in
the second case, the two ACKs will sometimes be only 100 ms apart.
(b) In the first case, TCP Vegas will measure RTT = 3 as soon as there is a
full window outstanding. This means ActualRate is down to 1 packet/sec.
However, BaseRTT is 1 sec, and so
ExpectedRate = CongestionWindow/BaseRTT is 3 packets/sec. Hence,
Diff is 2 packets/sec, and CongestionWindow will be decreased.

Chapter 6 91
In the second case, when a burst of three packets is sent the measured RTTs
are 1.0, 1.1, 1.2. Further measurements are similar. This likely does not re-
sult in enough change in the measured RTT to decrease ActualRate suffi-
ciently to trigger a decrease in CongestionWindow, and depending on the
value of α may even trigger an increase. At any rate, ActualRate decreases
much more slowly than in the first case.
44. If an application running over UDP has no congestion control, and it shares a
bottleneck link with an application that runs over congestion-controlled TCP,
then only the TCP traffic will reduce its sending rate in response to congestion.
In the extreme, the throughput of TCP traffic could drop to zero if there is enough
UDP traffic to congest the link on its own.
An application that receives RTCP receiver reports, however, can detect loss,
and attempt to emulate the congestion control behavior of TCP. For example, a
video application might respond to an RTCP report of packet loss by dropping
its transmission rate, perhaps by changing the video resolution or the strength of
its compression algorithm.
A detailed specification of a protocol that can run over UDP and yet respond to
congestion in a TCP-like way is in RFC 4340.
45. An ATM network may be only one network, or one type of network, in an inter-
net. Making service guarantees across such an ATM link does not in this setting
guarantee anything on an end-to-end basis. In other words, congestion manage-
ment is an end-to-end issue.
If IP operates exclusively over ATM, then congestion management at the ATM
level may indeed address total congestion (although if partial packet discard is
not implemented then dropped cells do not correspond very well to dropped
packets). In this setting, congestion control at the TCP level has the drawback
that it doesn’t address other protocols, and doesn’t take into account the switches’
knowledge of virtual circuits.
46. (a) Robot control is naturally realtime-intolerant: the robot can not wait indef-
initely for steering control if it is about to crash, and it can not afford to
lose messages such as “halt”, “set phasers on stun”, or even “switch to blue
paint”. Such an application could be adaptive in a setting where we have
the freedom to slow the robot down.
(b) If an application tolerates a loss rate of x, 0 x 1, then it is only receiv-
ing fraction 1 − x of the original bandwidth and can tolerate a reduction to
that bandwidth over a lossless link.
(c) Suppose the data being transmitted are positioning coordinates for some
kind of robotic device. The device must follow the positions plotted, though
some deviation is permitted. We can tolerate occasional lost data, by inter-
polating the correct path (there is a continuity assumption here); this would
qualify the application as loss-tolerant.
We also want to be able to claim that the application is non-adaptive. So
we will suppose that too much transmission delay means the robot cannot

Chapter 6 92
follow the path closely enough (or at least not with the required speed),
making the application non-delay-adaptive. A significant rate reduction,
similarly, might mean the device can’t keep to within the required tolerance
– perhaps it requires at least 80% of the coordinates – and so it would
qualify as non-rate-adaptive.
47. (a) One way to solve this is to imagine that we start with an empty bucket but
allow the bucket volume to become negative (while still providing pack-
ets); we then get the following table of bucket “indebtedness”: At T=0, for
example, we withdraw 8 tokens from the bucket (the number of packets
sent) and deposit 2 (the token rate).
Time, secs 0 1 2 3 4 5
Bucket volume -6 -8 -7 -5 -9 -8
We thus need an initial bucket depth of 9, so as not to run out at T=4.
Because all the volumes above are negative, the bucket with depth 9 never
overflows.
(b) If we do the same thing as above we get
Time, secs 0 1 2 3 4 5
Bucket volume -4 -4 -1 3 1 4
A bucket depth of 4 will thus accommodate T=0 and T=1. In this case
because the volume is sometimes positive we also need to check that the
bucket doesn’t overflow. If we start with an initially full bucket of depth 4,
we get
Time, secs 0 1 2 3 4 5
Bucket volume 0 0 2 4 2 4
Note that the bucket does become full of tokens at T=3 and T=5 but that
we are able to handle the 6 packets at T=4 because we had 4 tokens in the
bucket plus the 4 more tokens that arrive each interval. So 4 is the size of
the minimal token bucket.
49. (a) If the router queue is empty and all three flows dump their buckets at the
same time, the burst amounts to 15 packets for a maximum delay of 1.5 sec.
Since the router can keep up with packets due to steady-state traffic alone,
and can drain any earlier bucket dumps faster than the buckets get refilled,
such a burst is in fact the maximum queue.
(b) In 2.0 seconds the router can forward 20 packets. If flow1 sends an initial
burst of 10 at T=0 and another single packet at T=1, and flow2 sends 4 at
T=0 and 2 at T=1, that amounts to 17 packets in all. This leaves a minimum
capacity of 3 packets for flow3. Over the long term, of course, flow3 is
guaranteed an average of 8 packets per 2.0 seconds.
50. (a) If the router was initially combining both reserved and nonreserved traffic
into a single FIFO queue, then reserved flows before the loss were not get-
ting genuine service guarantees. After the loss the router is still handling all

Chapter 6 93
traffic via a single FIFO queue; the only difference is that all traffic is now
considered nonreserved. The state loss should thus make no difference.
(b) If the router used weighted fair queuing to segregate reserved traffic, then
a state loss may lead to considerable degradation in service, because the
reserved traffic now is forced to compete on an equal footing with hoi polloi
traffic.
(c) Suppose new reservations from some third parties reach the router before
the periodic refresh requests are received to renew the original reservations;
if these new reservations use up all the reservable capacity the router may
be forced to turn down the renewals.

1. Each string is preceded by a count of its length; the array of salaries is preceded
by a count of the number of elements. That leads to the following sequence of
integers and ASCII characters being sent:
7 R I C H A R D 4376 8 D E C E M B E R 2 1998 3 80000 85000 90000
2
5. Limited measurements suggest that, at least in one particular setting, use of htonl
slows the array-converting loop down by about a factor of two.
6. The following measurements were made on a 300MHz Intel system, compiling
with Microsoft’s Visual C++ 6.0 and optimizations turned off. We normalize to
the case of a loop that repeatedly assigns the same integer variable to another:
for (i=0;iN;i++) {j=k}
Replacing the loop body above with j=htonl(k) made the loop take about 2.9
times longer. The following homemade byte-swapping code took about 3.7 times
longer:
char * p = (char *) k;
char * q = (char *) j;
q[0]=p[3];
q[1]=p[2];
q[2]=p[1];
q[3]=p[0];
For comparison, replacing the loop body with an array copy A[i]=B[i] took about
2.8 times longer.
7. ASN.1 encodings are as follows:
INT 4 101
INT 4 10120
INT 4 16909060
9. Here are the encodings.
101 be 00000000 00000000 00000000 01100101
101 le 01100101 00000000 00000000 00000000
10120 be 00000000 00000000 00100111 10001000
10120 le 10001000 00100111 00000000 00000000
16909060 be 00000001 00000010 00000011 00000100
16909060 le 00000100 00000011 00000010 00000001
For more on big-endian versus little-endian we quote Jonathan Swift, writing in
Gulliver’s Travels:
94

...Which two mighty powers have, as I was going to tell you, been en-
gaged in a most obstinate war for six and thirty moons past. It began upon
the following occasion. It is allowed on all hands, that the primitive way of
breaking eggs before we eat them, was upon the larger end: but his present
Majesty’s grandfather, while he was a boy, going to eat an egg, and break-
ing it according to the ancient practice, happened to cut one of his fingers.
Whereupon the Emperor his father published an edict, commanding all his
subjects, upon great penalties, to break the smaller end of their eggs. The
people so highly resented this law, that our histories tell us there have been
six rebellions raised on that account.... Many hundred large volumes have
been published upon this controversy: but the books of the Big-Endians have
been long forbidden, and the whole party rendered incapable by law of hold-
ing employments.
11. The problem is that we don’t know whether the RPCVersion field is in big-
endian or little-endian format until after we extract it, but we need this informa-
tion to decide on which extraction to do.
It would be possible to work around this problem provided that among all the
version IDs assigned, the big-endian representation of one ID never happened to
be identical to the little-endian representation of another. This would be the case
if, for example, future versions of XDR continued to use big-endian format for
the RPCVersion field, but not necessarily elsewhere.
12. It is often possible to do a better job of compressing the data if one knows some-
thing about the type of the data. This applies even to lossless compression; it is
particularly true if lossy compression can be contemplated. Once encoded in a
message and handed to the encoding layer, all the data looks alike, and only a
generic, lossless compression algorithm can be applied.
13. [The DEC-20 was perhaps the best-known example of 36-bit architecture.]
Incoming 32-bit integers are no problem; neither are outbound character strings.
Outbound integers could either be sent as 64-bit integers, or else could lose the
high-order bits (with or without notification to the sender). For inbound strings,
one approach might be to strip them to 7 bits by default, make a flag available to
indicate whether any of the eighth bits had been set, and, if so, make available a
lossless mechanism (perhaps one byte per word) of re-reading the data.
14. Here is a C++ solution, in which we make netint⇒int an automatic conversion.
To avoid potential ambiguity, we make use of the explicit keyword in the con-
structor converting int to netint, so that this does not also become an automatic
conversion. (Note that the ambiguity would require additional code to realize.)
To support assignment netint = int, we introduce an assignment operator.
class netint {
public:
95

operator int() {return ntohl( netint);}
netint() : netint(0) // default constructor
explicit netint (int n) : netint(ntohl(n)) {}
netint operator=(int n) {
netint = htonl(n);
return *this;
}
int raw() {return netint;} // for testing
private:
int netint;
};
The above strategy doesn’t help at all with pointers, and not much with structures
and arrays. It doesn’t address alignment problems, for example.
15. Transmission bit order is the province of the network adapter, which addresses
this as it transmits or receives each byte. Generally all numeric formats on the
same machine (different sizes of ints, floats, etc) use the same bit order; only if
they didn’t would the programmer have to make distinctions.
16. For big-endian network byte order the average number of conversions is 0×p2
+
1 × 2p(1 − p) + 2 × (1 − p)2
. For receiver-makes-right this is 0 × p2
+ 1 ×
2p(1 − p) + 0 × (1 − p)2
; that is, if both sender and receiver are little-endian
then no conversion is done. These are evaluated below:
p = 0.1 p = 0.5 p = 0.9
big-endian network 1.80 1.00 0.20
receiver-makes-right 0.18 0.50 0.18
17. (a) Replace the markup tag text with corresponding codes. One or two bytes
would suffice for most XML languages.
(b) Represent numerical data using a numerical representation instead of text.
18. Try data files with lots of byte-string-level redundancy.
19. (a) letter encoding
a 1
b 01
c 001
d 000
(b) 1 × 0.5 + 2 × 0.3 + 3 × 0.1 + 3 × 0.1 = 1.7 So the compressed data uses
1.7/2 × 100 = 85% as many bits, or a 15% compression gain.
(c) The table is the same, although we could now give either a or b the 1-bit
encoding. The average compression is now 1 × 0.4 + 2 × 0.4 + 3 × 0.15 +
3 × 0.05 = 1.8, i.e., we use 90% as many bits or a 10% compression gain.
96

20. (a) This is a counting argument: there are 2N
strings of length N and only
20
+ 21
+ · · · + 2N−1
= 2N
− 1 strings of length N. Some string,
therefore, cannot get shorter.
(c) We let
c′
(s) =

0⌢
c(s) if length(c(s)) length(s)
1⌢
s otherwise
(where 0⌢
c(s) is c(s) with a zero-bit prepended). The initial bit is a flag
to indicate whether the remainder was compressed or not.
21. Bytes that occur with a run length of 1 we represent with themselves. If a byte
occurs in a run of more than 1, we represent it with the three bytes
[ESC] [count] [byte]
The byte [ESC] can be any agreed-upon escape character; if it occurs alone it
might be represented as [ESC][ESC].
22. A sample program appears on the web page; it generated the following data.
The uncompressed size of RFC 791 is 94892 bytes. There are 11,243 words
in all; the number of distinct words is 2255 and the dictionary size is 18226
bytes. The encoded size of the non-dictionary part with 12 bits per word is thus
(12×11243)/8 = 16865 bytes; together with the dictionary we get a compressed
size of 35091 bytes, 37% of the original size. There are 132 words appearing at
least 13 times, with a total frequency count of 6689. This means the 128 most
common words appear a total of 6637 times; this gives a compressed size of
(8 × 6637 + 13 × (11243 − 6637))/8 = 14122 bytes, plus the dictionary; the
total is now 34% of the original size. Note that exact numbers are sensitive to
the precise definition of a “word” used.
23. (a) For “symmetric” data such as this, coefficients DCT (i) for i = 1, 3, 5, 7
(starting the indexing at i = 0) should be zero or near-zero.
(b) If we keep six coefficients, the maximum error in pixel(i) after applying
the DCT and its inverse is about 0.7%, for i = 1 and i = 2. If we keep
only four or five coefficients (note that both choices lead to the same values
for the inverse DCT), then the maximum error is 6%, at i = 0; the error at
i = 1 is 5.6%.
(c) The input vectors for this problem look like h1, 0, 0, 0, 0, 0, 0, 0i with the 1
moving one position to the right as i increases. Here is a table listing, for
each si, the percentage error in the ith place of the final result. The smallest
error is for i = 0 and 7; the largest is for i = 1 and 6.
i 0 1 2 3 4 5 6 7
% error 12.3 53.1 39.6 45.0 45.0 39.6 53.1 12.3
24. The all-white image generates all zeros in the DCT phase. The quantization
phase leaves the 8×8 grid of 0’s unchanged; the encoding phase then compresses
it to almost nothing.
97

25. Here is the first row of an 8 ×8 pixmap consisting of a black line (value 0) in the
first column, and the rest all white:
00 FF FF FF FF FF FF FF
Here is the image after default-quality (cjpeg -quality 75) JPEG compression
and decompression; there is some faint vertical fringing (the columns with FC,
FD, and FE would appear as progressively fainter grey lines). All rows are iden-
tical; here is the first:
01 FC FF FD FF FF FE FF
With -quality 100, or even -quality 90, the fringing is gone; the image after
compression and decompression is identical to the original.
26. We start with an example specifically generated for the 8 × 8 grid; note that the
letters change gradually in both directions. Here is the original data:
a b c d e f g h
b c d e f g h i
c d e f g h i j
d e f g h i j k
e f g h i j k l
f g h i j k l m
g h i j k l m n
h i j k l m n o
We get the following after default-quality (quality=75) jpeg compression and
decompression; no letter is off by more than 1 ASCII value.
b b c d e g h h
b c d e f g h h
c d d f g h i i
d e f g h i j j
f f g h i j k l
g g h i j l l m
h h i j k l m n
h h i k l m n n
At quality=100 the text is preserved exactly. However, this is the best case.
Here is the first line of Lincoln’s Gettysburg Address,
Fourscore and seven years ago our fathers brought forth on this conti-
nent....,
compressed and decompressed. With spaces between words eliminated and ev-
erything made lowercase, at quality=75 we get:
hnruugdtdihjpkirmqicjlfgowpekoifappiosqrbnnjonkppqjioidjulafrnhq
At quality=100 we get:
fourscoreandsevenyeassagoovrfathersbroughtfnrthonthiscontinentan
The three errors are underlined. Leaving in the spaces, then at quality=100 we
get:
fourscpre and seven years ago our fathers bsought eosth on this
where the “ ” character is the character with decimal value 31, versus 32 for a
space character.
98

Chapter 7 99
Lowercase letters are all within the ASCII range 97-122 numerically. Space
characters are numerically 32; these thus appear to the DCT as striking disconti-
nuities.
27. Jpeg includes an encoding phase, but as this is lossless it doesn’t affect the image.
JPEG’s quantization phase, however, potentially rounds off all coefficients, to a
greater or lesser degree; the strategy here however leaves un-zeroed coefficients
alone.
28. Recalling that C(0) = 1/
√
2, we have
DCT (0, 0) =
1
2
√
2N
N−1
X
x=0
N−1
X
y=0
pixel(x, y)
=
N
√
N
2
√
2
×
1
N2
N−1
X
x=0
N−1
X
y=0
pixel(x, y)
=
N
√
N
2
√
2
× (average of the pixel(x, y)’s)
(2)
30. If you display I frames only, then the fast-forward speed is limited to the rate at
which I-frames are included; note that this may be variable.
The worst case for decoding an arbitrary frame is when the frame you want is
a B frame. It depends on a previous P frame P and a future P or I frame Q. To
decode the B frame you want, you will first need P and its I frame, and also Q.
If Q is a P frame, then its I frame is the same as that of P. The total number of
frames processed, including the one wanted, is four.
33. (a) For a while, the intervening B frames would show each macroblock con-
taining a point as an appropriately translated macroblock from the original
I frame. The δ for the frame is zero. Once the two points were close enough
that they were in the same macroblock, however, we would need a nonzero
δ to represent the frame, perhaps translating a macroblock from the original
I frame so as to show one point, and using a δ with one nonzero entry to
show the second point.
(b) If the points were of a fixed color, the only difference from the above is
that color macroblocks come at a different resolution. With points that are
changing in color, modest deltas are needed from the beginning to indicate
this.

2.
Plaintext Block 0
Encryption
Function
Plaintext Block 1
Plaintext Block 2
Plaintext Block 3
Initialization Vector
(For block 0 only)
Blocks of ciphertext
4. If the outputs are not truly random, then the algorithm becomes more vulnerable
to a brute force attack. For example, if some outputs are twice as likely to be
produced as others, an attacker can test fewer input strings to produce one of
those more likely outputs. A factor of two probably doesn’t matter too much, but
large non-randomness would weaken the hash significantly.
5. The adversary would replay, to Bob, the third message of the Needham-Schroeder
authentication protocol. Consistent with the protocol, Bob would extract the ap-
parently new, but actually old, session key, and use it to encrypt a nonce in a
reply. The adversary, having discovered the session key, would be able to de-
crypt the received nonce and reply as Alice would have.
6. We have password[N] = g(password[N − 1]); the essential property of g is
that it be believed that knowing g(x) does not provide any information that can
be used to find x.
7. (a) let q be the first N − 1 characters of the password p, of length N. The
eavesdropper is in possession of q at the point indicated in the hint; we now
assume that “sufficiently slowly” means that the eavesdropper can try all
passwords q⌢
ch, for all characters ch, before the original user has typed
101

the Nth character. If we assume passwords are restricted to 7-bit printable
ASCII, that’s only 96 tries.
(b) Other attacks include a compromised utility to calculate the one-time pass-
word f(mp, N) from the master password mp, discovery of a way to invert
the function g at least partially, eavesdropping on the initial selection of
mp, and “hijacking” a connection after authentication has been completed.
There are doubtless others, as well.
8. The FAQ at www.rsasecurity.com explains:
The Diffie-Hellman key exchange is vulnerable to a man-in-the-middle
attack. In this attack, an opponent Carol intercepts Alice’s public
value and sends her own public value to Bob. When Bob transmits
his public value, Carol substitutes it with her own and sends it to Al-
ice. Carol and Alice thus agree on one shared key and Carol and Bob
agree on another shared key. After this exchange, Carol simply de-
crypts any messages sent out by Alice or Bob, and then reads and
possibly modifies them before re-encrypting with the appropriate key
and transmitting them to the other party. This vulnerability is present
because Diffie-Hellman key exchange does not authenticate the par-
ticipants. Possible solutions include the use of digital signatures and
other protocol variants.
9. Because s is short, an exhaustive search conducted by generating all possible s
and comparing the MD5 checksums with m would be straightforward. Sending
MD5(s⌢
r), for some random or time-dependent r, would suffice to defeat this
search strategy, but note that now we would have to remember r and be able
to present it later to show we knew s. Using RSA to encrypt s⌢
r would be
better in that sense, because we could decrypt it at any time and verify s without
remembering r.
10. Each side chooses xi privately. They exchange signatures of their respective
choices as in the previous exercise, perhaps MD5(xi
⌢
ri) for random ri. Then
they exchange the actual xi’s (and ri’s); because of the signatures, whoever re-
veals their xi last is not able to change their choice based on knowing the other
xi. Then let x = x1 ⊕ x2; as long as either party chooses their xi randomly then
x is random.
11. Let PN be the probability that of N messages each checksum value is different
from all the preceding. As in Chapter 2 Exercise 41 we have
PN =

1 −
1
2128

1 −
2
2128

· · ·

1 −
N − 1
2128

Taking logs and approximating we get
logPN = −(1/2128
+ 2/2128
+ · · · + (N − 1)/2128
)
= −(1 + 2 + ... + (N − 1))/2128
≈ −N2
/2129
102

So PN ≈ e−N2
/2129
. For N = 263
the exponent here is −2126
/2129
= −1/8;
for N = 264
and N = 265
it is −1/2 and −2 respectively. Thus, the probabilities
are
P63 = e−1/8
= 0.8825,
P64 = e−1/2
= 0.6065,
P65 = e−2
= 0.1353.
The probability two messages have the same checksum is 1 − PN .
12. The problem with padding each 1-byte message with seven zero bytes before
encrypting is that we now are transmitting only 256 possible different encrypted
blocks and a codebreaking attack is quite straightforward.
Here are some better options. Each involves encrypting a full block for each
plaintext byte transmitted; the first two also require that we transmit a full block.
1. We could pad each plaintext byte with 7 random bytes before encrypting. This
is quite effective, if the random bytes are truly random.
2. We could make use of cipher block chaining, padding each plaintext byte
pi with seven zero-bytes before xoring with the previous Cipheri−1 block. A
roughly equivalent alternative, perhaps more like the previous option, is to pad
pi with seven bytes from Cipheri−1, and omit the xor.
3. So-called cipher-feedback (CFB) mode is sometimes used. Let ci denote the
ith encrypted byte. Given pi, we first use 3DES to encrypt the block hci−8 · · · ci−1i,
and then let ci be the xor of pi and the first byte of this encryption result. CFB
makes assumptions about the pseudo-randomness of 3DES output that may not
be initially apparent.
13. (a) Here is one possible approach. We add the following fields to the packets
(which presumably contain a Packet Type field and a Block Number field
already):
Sender’s time when the connection was initiated
Receiver’s time when the connection was initiated
Keyed MD5 checksum field
The latter consists of the MD5 checksum of everything else in the packet
concatenated with the sender’s own key. The packet sent does not include
the sender’s key. The recipient is able to recompute this checksum, because
we have assumed both keys are known to both parties.
The checksum field provides both message integrity and authentication; at
least it confirms that whoever created the packet knew the sender’s key.
The timestamps guard against replay attacks. Both sides must exchange
timestamps through a three-way handshake before any data is sent, much
like ISNs in TCP. If we include only the client’s timestamp, then the server
could only detect replay attacks by keeping track of the previous client
timestamp used. With both timestamps, each party is assured of a new
connection as long as it has chosen a new timestamp.
103

(b) The timestamps guard against late packets from a prior incarnation; older
incarnations would have at least one timestamp wrong. However, they do
nothing to protect against sequence number wraparound within a connec-
tion.
16. This is the chain-of-trust problem. Even though the root CA may have done
a fine job of checking the identity of the second-tier CA, it is hard for Bob to
know that the second-tier CA did the same for Alice. That is, Bob might doubt
whether the root CA checks the operations of the second-tier CA before signing
their certificate.
17. (a) The user gets a message the first time he or she attempts to connect to
the given server. At that point, the user can cancel, go ahead with the
connection this one time, or go ahead with the connection and save the
server’s public key for future authentications.
(b) Client authentication is up to the server. Password-based schemes are prob-
ably the most common. Public-key authentication is another possibility,
and PuTTY provides support for that.
(c) The user ranks the ciphers in order of preference, and PuTTY negotiates
with the server to use the user’s most preferred cipher among those sup-
ported by the server.
(d) Some servers may support only a weak cipher. PuTTY gives the user the
option of accepting the risk of using a weak cipher. The user sets a thresh-
old in the user’s ranking of the ciphers, and PuTTY warns the user if the
negotiated cipher falls below that threshold, giving the user the opportunity
to cancel the connection.
(e) The security derived from a session key decreases over time since more
time is available to perform cryptanalysis. It also decreases as more en-
crypted data becomes available for cryptanalysis. So a session key should
only be used short-term with a limited amount of data.
18. (a) An internal telnet client can presumably use any port ≥ 1024. In order for
such clients to be able to connect to the outside world, the firewall must
pass their return, inbound traffic. If the firewall bases its filtering solely
on port numbers, it must thus allow inbound TCP connections to any port
≥ 1024.
(b) If the firewall is allowed access to the TCP header Flags bits, then to block
all inbound TCP connections it suffices to disallow inbound packets with
SYN set but not ACK. This prevents an outside host initiating a connection
with an inside host, but allows any outbound connection. No port filtering
is needed.
19. (a) The FTP client uses the PORT command to tell the server what port to use
for the data connection. If this port can be taken from a limited range, in
which we are sure there are no other servers to which an outsider might
104

Chapter 8 105
attempt to connect, then a firewall can be configured to allow outside ac-
cess to this range (by examining both port numbers and the TCP Flags)
without unduly compromising security. This range cannot be limited to
a single port, though, because otherwise that port would frequently be in
TIME WAIT and unavailable.
(b) Instead of using the PORT command to tell the FTP server to what client
port it should connect for data transfer, an FTP client can send the server the
PASV (“passive”) command. The server response, assuming it supports
PASV, is to send an acknowledgment containing a server port number.
The client then initiates the data-transfer connection to this server port. As
this is typically an outbound connection through the firewall it can be safely
permitted.
20. The routers are configured as follows:
R1 blocks inbound traffic to the telnet port, unless the destination subnet is
net2.
R2 blocks all telnet traffic from net 2 to net 1.
21. The ISP might want to prohibit attacks (such as the IP spoofing attack described
in Exercise 5.17 or, for that matter, email spamming) launched by its own cus-
tomers.
22. RFC 2402 and RFC 2406 are handy for this exercise.
(a) IPsec ESP transport mode is incompatible with NAT. In the case of TCP/UDP
packets, NAT would need to update the checksum in TCP/UDP headers,
when an address in IP header is changed. However, as the TCP/UDP header
is encrypted by the ESP, NAT would not be able to make this checksum
update. As a result, TCP/UDP packets encrypted in transport mode ESP,
traversing a NAT device will fail the TCP/UDP checksum validation on the
receiving end and will simply not reach the target application.
(b) IPsec ESP tunnel mode may work with NAT. Since IPsec ESP tunnel mode
attach a new IP header and encapsulate the original IP packet in it. Since
the way ESP encryption works is that it only encrypts and authenticate the
IP payload, when the tunnel IP header gets stripped off, TCP/UDP check-
sum is preserved and still correct for the original IP packet. Therefore the
original IP packet can reach the target application.
(c) It is obvious that (a) will not work with PAT due to the same reasons above.
Now, as for the case (b), clearly IPsec ESP tunnel mode will not work with
PAT. PAT needs to look at the port numbers to do the translation, but those
are encrypted by ESP. The case (b) only works in true NAT case. There is
an effort called “NAT traversal” to work around this problem using UDP
encapsulation. With NAT traversal, the case (b) may work with PAT.

1. Both SMTP and HTTP are already largely organized as a series of requests sent
by the client, and attendant server reply messages. Some attention would have
to be paid in the request/reply protocol, though, to the fact that SMTP and HTTP
data messages can be quite large (though not so large that we can’t determine the
size before beginning transmission).
We might also need a MessageID field with each message, to identify which
request/reply pairs are part of the same transaction. This would be particularly
an issue for SMTP.
It would be quite straightforward for the request/reply transport protocol to sup-
port persistent connections: once one message was exchanged with another host,
the connection might persist until it was idle for some given interval of time.
Such a request/reply protocol might also include support for variable-sized mes-
sages, without using flag characters (CRLF) or application-specific size headers
or chunking into blocks. HTTP in particular currently includes the latter as an
application-layer issue.
3. Existing SMTP headers that help resist forgeries include mainly the Received:
header, which gives a list of the hosts through which the message has actually
passed, by IP address.
A mechanism to identify the specific user of the machine (as is provided by the
identd service), would also be beneficial.
4. If an SMTP host cannot understand a command, it responds with
500 Syntax error, command unrecognized
This has (or is supposed to have) no other untoward consequences for the con-
nection. A similar message is sent if a command parameter is not understood.
This allows communicating SMTPs to query each other as to whether certain
commands are understood, in a manner similar to the WILL/WONT protocol of,
say, telnet.
RFC 1869 documents a further mechanism: the client sends EHLO (Extended
HELO), and an EHLO-aware server responds with a list of SMTP extensions it
supports. One advantage of this is that it better supports command pipelining; it
avoids multiple exchanges for polling the other side about what it supports.
5. Further information on command pipelining can be found in RFC 2197.
(a) We could send the HELO, FROM, and TO all together, as these messages
are all small and the cost of unnecessary transmission is low, but it would
seem appropriate to examine the response for error indications before both-
ering to send the DATA.
(b) The idea here is that a server reading with gets() in this manner would be
unable to tell if two lines arrived together or separately. However, a TCP
106

buffer flush immediately after the first line was processed could wipe out
the second; one way this might occur is if the connection were handed off at
that point to a child process. Another possibility is that the server busyreads
after reading the first line but before sending back its response; a server
that willfully refused to accept pipelining might demand that this busyread
return 0 bytes. This is arguably beyond the scope of gets(), however.
(c) When the client sends its initial EHLO command (itself an extension of
HELO), a pipeline-safe server is supposed to respond with 250 PIPELIN-
ING, included in its list of supported SMTP extensions.
6. Implementers are free to add new subtypes to MIME, but certain default interpre-
tations may apply. For example, unrecognized subtypes of the application type
are to be treated as being equivalent to application/octet-stream. New experi-
mental types and subtypes can be introduced; names of such types are to begin
with X- to mark them as such. New image and text subtypes may be formally reg-
istered with the IANA; senders of such subtypes may also be encouraged to send
the data in one of the “standard” formats as well, using multipart/alternative.
7. We quote from RFC 1521:
NOTE: From an implementor’s perspective, it might seem more sen-
sible to reverse this ordering, and have the plainest alternative last.
However, placing the plainest alternative first is the friendliest pos-
sible option when multipart/alternative entities are viewed using a
non-MIME-conformant mail reader. While this approach does im-
pose some burden on conformant mail readers, interoperability with
older mail readers was deemed to be more important in this case.
It seems likely that anyone who has received MIME messages through text-based
non-MIME-aware mail readers would agree.
8. The base64 encoding actually defines 65 transmission characters; the 65th, “=”,
is used as a pad character. The data file is processed in input blocks of three bytes
at a time; each input block translates to an output block of four 6-bit pieces in
the base64 encoding process. If the final input block of the file contains one or
two bytes, then zero-bits are first added to bring the data to a 6-bit boundary (if
the final block is one byte, we add four zero bits; if the final block is two bytes,
we add two zero bits). The two or three resulting 6-bit pieces are then encoded
in the usual way, and two or one “=” characters are appended to bring the output
block to the required four pieces. In other words, if the encoded file ends with a
single =, then the original file size was ≡ 2 (mod 3); if the encoded file ends
with two =s then the original file size was ≡ 1 (mod 3).
9. (a) Enabling arbitrary SMTP relaying allows “spammers” to send unsolicited
email via someone else’s machine.
(b) One simple solution to this problem would be the addition of a password
option as part of the opening SMTP negotiation.
107

(c) One approach is to use a VPN, to make one’s external client IP address
appear to be internal. Relatively recently (relative to the length of time
SMTP has been around) RFC4954 specified “SMTP Service Extension for
Authentication”, which is in wide deployment today.
10. When the server initiates the close, then it is the server that must enter the TIME-
WAIT state. This requires the server to keep extra records; a server that averaged
100 connections per second would need to maintain about 6000 TIMEWAIT
records at any one moment. HTTP 1.1 has a variable-sized message transfer
mechanism; the size and endpoint of a message can be inferred from the head-
ers. The server can thus transfer a file and wait for the client to detect the end
and close the connection. Any request-reply protocol that could be adapted to
support arbitrarily large messages would also suffice here.
11. For supplying an alternative error page, consult www.apache.org or the docu-
mentation for almost any other web server; apache provides a setting for Error-
Document in httpd.conf.
RFC 2068 (on HTTP) states:
10.4.5 404 Not Found
The server has not found anything matching the Request-URI [Uni-
form Resource Identifier, a more general form of URL].
However, nothing in RFC 2068 requires that the part of a URL following the host
name be interpreted as a file name. In other words, HTTP servers are allowed to
interpret “matching”, as used above, in whatever manner they wish; in particular,
a string representing the name of a nonexistent file may be said to “match” a
designated ErrorDocument. Another example of a URL that does not represent
a filename is a dynamic query.
12. One server may support multiple web sites with multiple hostnames, a technique
known as virtual hosting. HTTP GET requests are referred by the server to the
appropriate directory based on the hostname contained in the request.
13. A TCP endpoint can abort the connection, which entails the sending of a RST
packet rather than a FIN. The endpoint then moves directly to TIMEWAIT.
To abort a connection using the Berkeley socket library, one first sets the SO LINGER
socket option, with a linger time of 0. At this point an application close() trig-
gers an abort as above, rather than the sending of a FIN.
14. (a) A mechanism within HTTP would of course require that the client browser
be aware of the mechanism. The client could ask the primary server if
there were alternate servers, and then choose one of them. Or the primary
server might tell the client what alternate to use. The parties involved might
measure “closeness” in terms of RTT, in terms of measured throughput, or
(less conveniently) in terms of preconfigured geographical information.
108

(b) Within DNS, one might add a WEB record that returned multiple server
addresses. The client resolver library call (e.g. gethostbyname()) would
choose the “closest”, determined as above, and return the single closest
entry to the client application as if it were an A record. Also, some CDNs
today use DNS resolution to try to direct a client to a nearby CDN node.
This is usually done without the client’s knowledge.
See the answer to question 37 for more on this topic.
15. The number of B2B and EAI network applications is potentially huge. A Web
Services protocol framework simplifies the task of specifying, developing, and
maintaining their protocols, and managing their operation. (Imagine creating
and running, say, one million custom FTP- and SMTP-like protocols!)
16. Amazon’s S3 (Simple Storage Service) Web Service is a fee-based, high-availability,
high-speed Internet storage service. Its storage model has “buckets” that are like
directories; they contain “objects.”
The SOAP operations are:
• ListAllMyBuckets
• CreateBucket
• DeleteBucket
• ListBucket
• GetBucketAccessControlPolicy
• SetBucketAccessControlPolicy
• PutObjectInline
• PutObject
• GetObject
• GetObjectExtended
• DeleteObject
• GetObjectAccessControlPolicy
• SetObjectAccessControlPolicy
The REST operations are HTTP operations. Their interpretation depends on
what strings are appended to the base URI of the web service. For example,
the URI http://s3.amazonaws.com/foo refers to the bucket “foo,” while the URI
http://s3.amazonaws.com/foo/bar refers to the object “bar” in the bucket “foo.”
The base URI http://s3.amazonaws.com may be considered to refer to the user’s
overall S3 account (both SOAP and REST S3 APIs authenticate the sender of a
request).
Using this convention, the REST operations (and their equivalents in parenthe-
ses) are:
• GET Service (ListAllMyBuckets)
109

• PUT Bucket (CreateBucket)
• GET Bucket (ListBucket)
• DELETE Bucket (DeleteBucket)
• PUT Object (PutObject, PutObjectInline). The SOAP interface gives you a
choice between transmitting the Object in the body of the SOAP message,
or as a DIME attachment.
• GET Object (GetObject, GetObjectExtended). GetObjectExtended sup-
ports some conditionals similar to those provided by HTTP GET headers,
e.g. return the object only if it has been modified since a specified time.
• HEAD Object. Retrieves metadata about the object.
• DELETE Object (DeleteObject)
The REST interface does not have distinct operations for access control policies.
“There are two ways to set the access control policy with REST. You can set the
access control list (ACL) for an existing bucket or object by requesting a PUT to
/bucket?acl or /bucket/key?acl. Or, at the time you are writing a bucket or object
you can include an x-amz-acl header with your PUT request that stores a canned
ACL with the written resource.”
Note that the resource/service in this example, with its buckets and objects, maps
nicely onto URIs. Other resources/services might have to embed comparable
information in the data instead of the URI, which would probably be more in
line with the REST philosophy.
17. Consider the GetObject operation of Amazon’s S3 (Simple Storage Service) Web
Service.
GetObject’s input message is a GetObjectRequest, and its output message is a
GetObjectResponse:
wsdl:operation name=GetObject
wsdlsoap:operation soapAction=/
wsdl:input name=GetObjectRequest
wsdlsoap:body use=literal/
/wsdl:input
wsdl:output name=GetObjectResponse
wsdlsoap:body use=literal/
/wsdl:output
/wsdl:operation
A GetObjectRequest message consists of a GetObject element
wsdl:message name=GetObjectRequest
wsdl:part element=tns:GetObject name=parameters/
/wsdl:message
110

which has the following fields:
xsd:element name=GetObject
xsd:complexType
xsd:sequence
xsd:element name=Bucket type=xsd:string/
xsd:element name=Key type=xsd:string/
xsd:element name=GetMetadata type=xsd:boolean/
xsd:element name=GetData type=xsd:boolean/
xsd:element name=InlineData type=xsd:boolean/
xsd:element name=AWSAccessKeyId type=xsd:string minOccurs=0/
xsd:element name=Timestamp type=xsd:dateTime minOccurs=0/
xsd:element name=Signature type=xsd:string minOccurs=0/
xsd:element name=Credential type=xsd:string minOccurs=0/
/xsd:sequence
/xsd:complexType
/xsd:element
A GetObjectResponse message consists of a GetObjectResponse element
wsdl:message name=GetObjectResponse
wsdl:part element=tns:GetObjectResponse name=parameters/
/wsdl:message
of type GetObjectResult
xsd:element name=GetObjectResponse
xsd:complexType
xsd:sequence
xsd:element name=GetObjectResponse type=tns:GetObjectResult/
/xsd:sequence
/xsd:complexType
/xsd:element
which has the following fields:
xsd:complexType name=GetObjectResult
xsd:complexContent
xsd:extension base=tns:Result
xsd:sequence
xsd:element name=Metadata
type=tns:MetadataEntry minOccurs=0 maxOccurs=unbounded/
111

xsd:element name=Data
type=xsd:base64Binary nillable=true/
xsd:element name=LastModified type=xsd:dateTime/
xsd:element name=ETag type=xsd:string/
/xsd:sequence
/xsd:extension
/xsd:complexContent
/xsd:complexType
18. One option would be to allow some receivers to join only the audio portion of the
conference while others join both audio and video; these could be advertised as
separate sessions. Or there could be high bandwidth and low bandwidth sessions,
with the speaker sending to all sessions but receivers joining the session that suits
their bandwidth. SAP would be used to notify the receivers of what types of
session are available.
A second approach would be to send all the media to a central “mixer” which
could then transmit a single audio stream representing the audio mix of all cur-
rent speakers and a single video stream, perhaps showing the face of the current
speaker or speakers.
19. For audio data we might send sample[n] for odd n in the first packet, and for even
n in the second. For video, the first packet might contain sample[i,j] for i+j odd
and the second for i+j even; dithering would be used to reconstruct the missing
sample[i,j] if only one packet arrived.
JPEG-type encoding (for either audio or video) could still be used on each of
the odd/even sets of data; however, each set of data would separately contain the
least-compressible low-frequency information. Because of this redundancy, we
would expect that the total compressed size of the two odd/even sets would be
significantly larger than what would be obtained by conventional JPEG compres-
sion of the data.
20. URI includes URL and URN. URN (Uniform Resource Name) is a persistent
and location-independent name in a namespace, while URL specifies how to
“locate” the resource. So, a URL can be moved or disappear, but a URN cannot.
An example of URN is urn:ISBN:0-201-62433-8, which refers to a book using
the ISBN namespace.
21. MX records supply a list of hosts able to receive email; each listed host has an
associated numeric “mail preference” value. This is documented further in RFC
974. Delivery to the host with the lowest-numbered mail preference value is to
be attempted first.
For HTTP, the same idea of supporting multiple equivalent servers with a single
DNS name might be quite useful for load-sharing among a cluster of servers;
however, one would have to ensure that the servers were in fact truly stateless.
Another possibility would be for a WEB query to return a list of HTTP servers
each with some associated “cost” information (perhaps related to geographical
distance); a client would prefer the server with the lowest cost.
112

22. ARP traffic is always local, so ARP retransmissions are confined to a small area.
Subnet broadcasts every few minutes are not a major issue either in terms of
bandwidth or CPU, so a small cache lifetime does not create an undue burden.
Much of DNS traffic is nonlocal; limiting such traffic becomes more important
for congestion reasons alone. There is also a sizable total CPU-time burden on
the root nameservers. And an active web session can easily generate many more
DNS queries than ARP queries. Finally, DNS provides a method of including
the cache lifetime in the DNS zone files. This allows a short cache lifetime to be
used when necessary, and a longer lifetime to be used more commonly.
If the DNS cache-entry lifetime is too long, however, then when a host’s IP
address changes the host is effectively unavailable for a prolonged interval.
23. DNS servers will now take on ARP’s role, by in effect supplying both subnet
number and physical address of hosts in its domain. DNS servers must therefore
now monitor hosts for possibly changed physical addresses.
A fairly common method in IPv4 of finding ones DNS server is via static con-
figuration, e.g. the Unix /etc/resolv.conf files. If this mechanism were still used
in IPv6, changing the Ethernet address of a local DNS server would now involve
considerable updating, both of the local clients and also the DNS parent. IPv6
clients, however, are likely to find their DNS server dynamically, e.g. via DHCP,
instead.
24. The lookup method here requires trusting of the remote site’s DNS PTR data,
which may not be trustworthy. Suppose, for example, that it is known that ci-
cada.cs.princeton.edu trusts gnat.cs.princeton.edu. A request for authen-
tication might arrive at cicada from, say, IP address 147.126.1.15, which is
not part of the princeton.edu domain. If cicada followed the strategy of the
exercise here, it would look up the string 15.1.126.147.in-addr.arpa in the
DNS PTR data. This query would eventually reach the DNS server for PTR
zone 1.126.147.in-addr.arpa, which if suborned or malicious might well return
the string gnat.cs.princeton.edu regardless of the fact that it had no connec-
tion with princeton.edu. Hostname strings returned by DNS servers for PTR
searches are arbitrary, and need not be related to the server’s assigned domain
name.
A forward DNS lookup to confirm the result of the reverse DNS lookup would,
however, be reasonably safe.
25. There is little if any relationship, formally, between a domain and an IP network,
although it is nonetheless fairly common for an organization (or department) to
have its DNS server resolve names for all the hosts in its network (or subnet),
and no others. The DNS server for cs.princeton.edu could, however, be on
a different network entirely (or even on a different continent) from the hosts
whose names it resolves. Alternatively, each x.cs.princeton.edu host could
be on a different network, and each host that is on the same network as the
cs.princeton.edu nameserver could be in a different DNS domain.
113

Chapter 9 114
If the reverse-mapping PTR records are used, however, then the same name-
server can handle both forward and reverse lookups only when DNS zones do
correspond to groups of subnets.
26. If a host uses a nonlocal nameserver, then the host’s queries don’t go into the
local nameserver’s cache (although this is only relevant if there is some reason
to believe some other local host might make use of the cached entries). Queries
have farther to travel, too. Otherwise there is no penalty for having the “wrong”
DNS server.
The DNS traffic volume will be the same for the nonlocal nameserver as for a
local nameserver, if the nonlocal nameserver is “on the way” to the nameserver
that ultimately holds the address requested. Use of a nonlocal nameserver could
result in less DNS traffic if the nonlocal nameserver has the entries in its cache,
and isn’t too far away, but local nameserver does not. This might be the case
if, for example, a large group of people with similar interests all used the same
nonlocal nameserver.
27. Figure 9.17 is “really” a picture of the domain hierarchy again. Nameservers
have been abstracted, effectively, into one per zone (duplicates are consolidated,
and a nameserver serving multiple zones would appear in multiple entries).
Without this abstraction, a graph of all nameservers would simply be all DNS
servers joined by edges corresponding to NS records, from zone parent to child.
It would not necessarily be acyclic, even as a directed graph.
28. Here is an example based on princeton.edu. whois princeton.edu returns:
Domain Name: PRINCETON.EDU
Registrant:
Princeton University
Office of Information Technology
701 Carnegie Center, Suite 302
Princeton, NJ 08540
UNITED STATES
...
29. One would first look up the IP address of the web server, using, say, host or
dig. One would then use whois to look up who is assigned that IP address, and
compare the resulting identification to that obtained by using whois to look up
the web server domain name.
30. (a) One could organize DNS names geographically (this hierarchy exists al-
ready; chi.il.us is the zone for many sites in the Chicago area), or else
organize by topic or service or product type. The problems with these alter-
natives are that they tend to be harder to remember, and there is no natural

Chapter 9 115
classification for corporations. Geography doesn’t work as large corpora-
tions are not localized geographically. Classifying by service or product
has also never been successful; this changes too quickly as corporations
merge or enter new areas or leave old ones.
(b) With multiple levels there are lots more individual nameserver queries, and
the levels are typically harder to remember.
31. If we just move the .com entries to the root nameserver, things wouldn’t be much
different than they are now, in practice. In theory, the root nameservers now
could refer all queries about the .com zone to a set of .com-specific servers; in
practice the root nameservers (x.root-servers.net for x from a to m) all do an-
swer .com queries directly. (They do not, however, answer .int queries directly.)
The proposal here simply makes this current practice mandatory, and shouldn’t
thus affect current traffic at all, although it might leave other zones such as .org
and .net and .edu with poorer service someday in the future.
The main problem with moving the host-level entries, such as for www.cisco,
to a single root nameserver entry such as cisco, is that this either limits orga-
nizations to a single externally visible host, or else (if the change is interpreted
slightly differently) significantly increases root nameserver traffic as it returns
some kind of block of multiple host addresses. In effect this takes DNS back to
a single central server. Perhaps just as importantly, the updating of the IP ad-
dresses corresponding to host names is now out of the hands of the organizations
owning the host names, leading to a considerable administrative bottleneck.
However, if we’re just browsing the web and need only one address for each
organization, the traffic would be roughly equivalent to the way DNS works
now. (We are assuming that local resolvers still exist and still maintain request
caches; the loss of local caches would put an intolerable burden on the root
nameservers.)
32. DNS records contain a TTL value, specified by the DNS server, representing
how long a DNS record may be kept in the client cache. RFC 1034 puts it this
way:
If a change can be anticipated, the TTL can be reduced prior to the
change to minimize inconsistency during the change, and then in-
creased back to its former value following the change.
33. Here is a series of dig queries and edited responses. First we try to find top level
servers for the edu. domain:
% dig +norecurse edu.
;; AUTHORITY SECTION:
edu. 143993 IN NS C3.NSTLD.COM.
edu. 143993 IN NS D3.NSTLD.COM.

Chapter 9 116
edu. 143993 IN NS E3.NSTLD.COM.
edu. 143993 IN NS G3.NSTLD.COM.
...
Next, picking one of the top level servers:
% dig @c3.nstld.com princeton.edu.
princeton.edu. 172800 IN NS NS1.UCSC.edu.
princeton.edu. 172800 IN NS NS2.FAST.NET.
princeton.edu. 172800 IN NS NS3.NIC.FR.
princeton.edu. 172800 IN NS DNS.princeton.edu.
;; ADDITIONAL SECTION:
DNS.princeton.edu. 172800 IN A 128.112.129.15
NS1.UCSC.edu. 172800 IN A 128.114.142.6
Now we can query the name server for Princeton:
% dig @128.112.129.15 cs.princeton.edu
cs.princeton.edu. 172800 IN NS ns2.fast.net.
cs.princeton.edu. 172800 IN NS ns3.fast.net.
cs.princeton.edu. 172800 IN NS dns1.cs.princeton.edu.
;; ADDITIONAL SECTION:
ns1.fast.net. 62914 IN A 209.92.1.12
ns1.ucsc.edu. 43200 IN A 128.114.142.6
ns2.fast.net. 62914 IN A 206.245.170.12
dns1.cs.princeton.edu. 172800 IN A 128.112.136.10
dns2.cs.princeton.edu. 172800 IN A 128.112.136.12
And then on to one of the CS department’s servers:
% dig @128.112.136.12 www.cs.princeton.edu
;; ANSWER SECTION:
www.cs.princeton.edu. 21600 IN CNAME coreweb.cs.princeton.edu.
coreweb.cs.princeton.edu.
21600 IN A 128.112.136.35

Chapter 9 117
34. (b) Use the name of each object returned as the snmpgetnext argument in the
subsequent call.
35. For example, you can alternate SNMP queries with telnet connections to an oth-
erwise idle machine, and watch tcp.tcpPassiveOpens and tcp.tcpInSegs tick
up appropriately. One can also watch tcp.tcpOutSegs.
36. By polling the host’s SNMP server, one could find out what rsh connections
had been initiated. A host that receives many such connections might be a good
candidate for attack, although finding out the hosts doing the connecting would
still require some guesswork.
Someone able to use SNMP to set a host’s routing tables or ARP tables, etc,
would have many more opportunities.
37. A CDN using only HTTP 302 redirects could operate as follows: the user points
his browser at the origin, say http://www.example.com/page.html, and the ori-
gin redirects using the HTTP 302 response to tell the browser to talk to another
node that is part of the CDN, say, http://www.cdn.com/www.example.com/
page.html, and that node, using whatever algorithms the CDN implements to
choose a suitable surrogate, issues another HTTP 302 redirect to the appropriate
node with the content.
The main limitations of this approach are that, as described, there is still some
load on the origin to issue the redirect, and there is added latency to process each
redirect.
A CDN using only DNS can (with the permission of the content owner) establish
a DNS CNAME for www.example.com,so that DNS queries to www.example.com
would be translated to queries to some other domain under the control of the
CDN operator, say a123.cdn.com. The CDN operator can then resolve queries to
the DNS for a123.cdn.comto a suitable surrogate for the content at www.example.
com.
The main limitation of this approach is that it limits the granularity of infor-
mation that is available to the CDN for making its choice of surrogates; as de-
scribed, all the content from www.example.com would have to be served by the
same surrogates. You could use multiple domains (e.g. images.example.com,
videos.example.com) to address this issue.
Another limitation of using DNS is that the CDN doesn’t know as much about
the client (e.g. its IP address) because the DNS resolver operated by the CDN
operator gets DNS queries from other resolvers, not the clients. This can limit
the CDN’s ability to pick a surrogate that is close to the client, for example.
A combination is possible. For example, you could use DNS to get www.example.
com to redirect to a123.cdn.com, and then use HTTP 302 redirects to pick a suit-
able surrogate based on the content requested. Note that this overcomes most of
the prior limitations.
38. One problem would be the caching of DNS responses; hosts would tend to keep
going to the overloaded server which their local DNS cache considers to be the

Chapter 9 118
right choice, unless every DNS server in the hierarchy implements the redirection
scheme.
To circumvent that problem, TTL should be set to a small number, and this causes
more queries to the DNS servers higher up the hierarchy, creating potentially
high loads on those servers.
39. The following describes one possible approach; many others are possible.
We build on the answer to question 37. Assume that a content origin such as
www.example.com is directed by means of a DNS CNAME to a node operated
by CDN A, which in turn uses HTTP 302 redirects to direct a browser to a
surrogate. We can imagine that both CDNs have some designated machines
that are responsible for communicating with other CDNs. One such machine in
CDN A tells an equivalent machine in CDN B (using some protocol on which
they agree) about
• the set of content that CDN A knows how to obtain (such as the content
from www.example.com), and
• the set of IP address prefixes that CDN A considers to correspond to its end
users.
CDN B also provides the corresponding information about its content and end-
user prefixes to CDN A. Having exchanged this information, the two CDNs make
use of it in their routing of requests to surrogates.
Consider a request from a client that is among the prefixes served by CDN B
for a piece of content from www.example.com. Thanks to the DNS CNAME,
that request will be handled by a node operated by CDN A. But since CDN B
provided its list of end user prefixes to CDN A, CDN A can use an HTTP 302
redirect to send the request over to CDN B. CDN B can use its normal methods to
redirect the request to a suitable surrogate in CDN B. In addition, CDN B needs
to be able to actually obtain the content from www.example.com at some point
in order to cache it in a surrogate. This might be done by retrieving the content
from a node in CDN A, or by going back to www.example.com. The former
might be preferred since it allows example.com to have a relationship with only
a single CDN.
40. If a cache node has high cache miss rates, then adding more storage to that node
should help to improve that aspect of performance. In particular, one would look
for cache misses that arise because a lack of capacity has forced the eviction of
content that was later requested.
High CPU load or contention for disk access at the root or intermediate layers
of the hierarchy would be a sign that the cost of servicing requests from the
lower layers is getting too high, which might imply that adding another level of
hierarchy would help.
41. Suppose many people want to watch a popular TV show at the same time. A
multicast overlay could be used to stream the video packets from a single origin

Chapter 9 119
to all the viewers. The nodes in the overlay from a tree and the viewers are the
leaves.
Now suppose those viewers all want to watch the show, but at different times.
They could each subscribe to the multicast feed of video packets, and write the
packets to a local disk for later viewing, rather like a digital video recorder such
as TiVo. Alternatively, a copy of the show could be placed in a CDN. Assume
the CDN is arranged as a hierarchy, as in question 40. Rather than storing the TV
show at home to watch later, each viewer can request the show from the CDN
when he wants to watch it. A leaf node in the CDN that doesn’t already have the
show will fetch it from a node higher up in the hierarchy; these requests flow up
to the root of the hierarchy. So in both cases, a tree of overlay nodes distributes
a single copy from the root to the leaves, the only difference being whether the
content is pulled by the leaves or pushed from the root and stored at the leaves.
42. (a) All peers could have the file after n time units. During each time unit, each
peer with the piece can transmit it to one peer without the piece, so the
number of peers with the piece doubles with each time unit: 1 (= 20
) at
time 0, 2 (= 21
) at time 1, 4 (= 22
) at time 2, up to 2n
peers at time n.
(b) All peers could have the file after less than 2n time units. If all pieces were
downloaded to just the right peers in just the right order, it is possible for
all peers to obtain the file after as few as n + 2 time units.
Let’s label the two pieces A and B. Let’s label as PA the peer that initially
has the file. During the first time unit, PA transmits B to another peer, call
that other peer PB. Split the peers into two equal groups of 2(n−1)
, one
containing PA and the other containing PB. Now, from the result of the first
question, we know that all the peers grouped with PA can obtain A within
an additional n − 1 time units. Because the two sets of peers are disjoint,
with no interference or contention between them, all the peers grouped with
PB can obtain B during the same n − 1 time units.
Together with the initial step, we have used n time units so far. Another
time unit will suffice for the peers grouped with PA to transmit A to all
the peers grouped with PB. One more time unit will suffice for the peers
grouped with PB to transmit B to all the peers grouped with PA (except PA
itself, which has had B from the beginning). The 2 time units required for
each half to transmit its piece to the other half increases the total to n + 2
time units.
It can also be shown that n + 2 is the minimum (assuming n is at least 3).
We know that it must take exactly (2n
) − 1 individual upload/download
transactions to propagate A from PA (directly or indirectly) to all of PA’s
(2n
) − 1 peers. Another (2n
) − 1 transactions is required for B, making a
total of 2(n+1)
− 2 transactions.
On the other hand, not enough peers have pieces to all participate in trans-
actions during the first n−1 time units. During that period, an upper bound
on the number of transactions that can occur during the interval t − 1 to t

Chapter 9 120
to 2(t−1)
. Some arithmetic gives an upper bound of 2(n−1)
−1 transactions
total during the first n − 1 time units.
For an upper bound on the transactions during each subsequent time unit,
let’s assume every peer is able to participate in a transaction. Then the
number of transactions would be 2(n−1)
per time unit. So the two time units
from n−1 to n+1 add 2n
transactions to the previous upper bound. Hence
an upper bound on the number of transactions that could occur during the
first n + 1 time units is 2n
+ 2(n−1)
− 1. Some arithmetic shows that, for n
at least 3, this is less than the 2(n+1)
− 2 transactions required. Thus n + 2
is the minimum time required.

Text book 2 computer networks_a systems approach_peterson solution manual

More Related Content

What's hot (20)

Similar to Text book 2 computer networks_a systems approach_peterson solution manual (20)

Recently uploaded (20)

Text book 2 computer networks_a systems approach_peterson solution manual