Delay jitter control for real time communication

In Proceedings of TriComm ’91
Delay Jitter Control for Real-Time Communication
in a Packet Switching Network
Dinesh C. Verma
Hui Zhang
Domenico Ferrari
Computer Science Division
Department of Electrical Engineering and Computer Sciences
University of California at Berkeley
& International Computer Science Institute
Berkeley, California
ABSTRACT
A real-time channel is a simplex connection
between two nodes characterized by parameters
representing the performance requirements of the
client. These parameters may include a bound on the
minimum connection bandwidth, a bound on the max-
imum packet delay, and a bound on the maximum
packet loss rate. Such a connection may be esta-
blished in a packet-switching environment by means
of the schemes described by some of the authors in
previous papers. In this paper, we study the feasibil-
ity of bounding the delay jitter for real-time channels
in a packet-switched store-and-forward wide-area net-
work with general topology, extending the scheme
proposed in the previous papers. We prove the
correctness of our solution, and study its effectiveness
by means of simulations. The results show that the
scheme is capable of providing a significant reduction
in delay jitter, that there is no accumulation of jitter
along the path of a channel, and that jitter control
reduces the buffer space required in the network
significantly.
KEYWORDS: Real-time communication,
delay jitter, packet-switching network, real-
time channel.
INTRODUCTION
Real-time communication services will become
a necessity in broadband integrated networks, espe-
cially if digital audio and digital video attain the
prominence being predicted for them. A real-time
communication service allows a client to transport
information with performance guarantees. The
specific performance guarantees that will be needed
will depend on the type of traffic (see [Ferrari 90] for
a discussion of these requirements). It is likely that
other kinds of traffic (i.e., other than audio or video)
will also like to take advantage of guaranteed perfor-
mance communication. We feel that real-time com-
munication services should be an integral part of
future integrated networks, coexisting with the tradi-
tional connectionless and connection-oriented ser-
vices provided by present communication networks
[Leiner 89].
The schemes to provide real-time communica-
tion can be broadly categorized under the following
three switching techniques: circuit switching, packet
switching, and hybrid switching. An integrated net-
work based on either circuit switching or hybrid
switching typically has very poor resource utilization
when bursty traffic needs to be provided with perfor-
mance guarantees. In addition, hybrid switching
requires more complex switches, and does not con-
form to the goal of fully integrated networks. Full
integration is more likely to be achieved by packet
switching. However, while packet switching can pro-
¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡
This research has been supported in part by AT&T Bell La-
boratories, Hitachi, Ltd., the University of California under a
MICRO grant, the National Science Foundation under Grant
No. CDA-8722788, and the International Computer Science
Institute. The views and conclusions in this document are
those of the authors and should not be interpreted as
representing official policies, either expressed or implied, of
any of the sponsoring organizations.

vide performance guarantees regarding delays or loss
rates (see [Ferrari 89], [Ferrari 90a] and [Ferrari 90b]
for such schemes), it is not very convenient for traffic
requiring low delay variation or jitter.
A bound on delay jitter is required by both
interactive and non-interactive applications involving
digital continuous media to achieve an acceptable
quality of sound and animated images. Delay jitter
can be eliminated by buffering at the receiver. How-
ever, the amount of buffer space required at the
receiver can be reduced if the network can provide
some guarantees about delay jitter as well. The reduc-
tion can be significant for high bandwidth communi-
cation. It therefore makes sense to ask whether the
schemes providing bounded delays and loss rates can
be extended to provide any kind of delay jitter
guarantees, and, if so, under what conditions and at
what cost. As it turns out, the mechanism to reduce
jitter reduces the amount of buffer space required not
only in the receiver but also within the network.
From the point of view of a client requiring
bounded delay jitter, the ideal network would look
like a constant delay line, where packets handed to the
network by the sending entity are given to the receiv-
ing entity after a fixed amount of time. The jitter of a
connection can thus be defined by the maximum abso-
lute difference in the delays experienced by any two
packets on that connection. 1 In conjunction with a
bound on the maximum delay, a delay jitter guarantee
enforces both the maximum and the minimum delay
to be experienced by a packet on the channel. The
goal of the delay jitter control algorithm, to be
described below, is to keep the delay experienced by
any packet on a connection within these two bounds,
which are specified at connection establishment time.
This paper describes a method for guaranteeing
delay jitter in a packet-switching wide-area network,
and presents an evaluation of some of its most impor-
tant characteristics. Since it is an extension to an
existing scheme [Ferrari 90a], the method can be used
in all environments in which the original scheme can
be used. However, like the original scheme, we will
present it in the context of a contemporary
connection-oriented packet-switching store-and-
forward network, and evaluate it by simulation in the
same context. Thus, we assume that our network can
be modeled as a mesh of nodes connected by links
with constant propagation time. Links which do not
have a constant propagation time should provide a
bound on the maximum delay jitter they can intro-
duce.
The paper is organized as follows: we first
revisit the original scheme that provides delay
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
1 Although a bound on the end-to-end delay of a real-time
channel is a bound on the delay jitter as well, it is too loose to
be acceptable.
bounds, and then sketch the modifications required to
add the delay jitter bounds to the existing scheme.
We subsequently describe the simulation experiments
we ran to evaluate our scheme, and discuss the results
we obtained. Finally, we draw our conclusions.
THE ORIGINAL SCHEME
Real-time communication, as envisioned in
[Ferrari 90a], is based on simplex fixed-route connec-
tions to be called real-time channels or simply chan-
nels, whose routes will be chosen at establishment
time. In order to provide real-time service, clients are
required to declare their traffic characteristics and
performance requirements 2 at the time of channel
establishment according to the following parameters:
£
for the offered load
¤
the minimum packet interarrival time on
the channel, xmin,
¥
the maximum packet size smax, and
¦
the maximum service time t in the node
for the channel’s packets; this includes
the time required for transmission, header
processing, and any other operations the
node may need to perform for the packet;
§
for the performance bounds
¨
the source-to-destination delay bound D
for the channel’s packets.
For simplicity, we assume that channels require
that there be no packet loss due to buffer overruns in
any of the intermediate nodes, and that the client is
able to tolerate losses due to the other sources of error
in the network.
The original scheme consists of three parts: an
establishment procedure, a scheduling mechanism,
and a rate control mechanism.
The Establishment Procedure
The channel establishment mechanism may be
built on top of any procedure that can be used to set
up connections. The goal of the establishment pro-
cedure is to break up the end-to-end delay bound Di
required by channel i into the local delay bounds di, j
at each intermediate node j. The local bounds are
computed so that, if a node j can assure that no packet
©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©¡©
2 We state here only those traffic and performance parame-
ters of the original method that are required to provide a
bound on delay jitter. Thus, the parameters mentioned restrict
the original scheme to providing deterministic delay bounds
for smooth traffic, the kind most likely to require a bound on
delay jitter. It is possible to extend the original scheme with
its full set of parameters to provide a probabilistic delay jitter
guarantee and to incorporate bursty traffic, but we restrict our-
selves to the simpler case to keep the length of this paper
within reasonable bounds.

on channel i will be delayed locally beyond its local
bound di, j, the end-to-end delay bound Di can be
met. As the establishment request moves from the
source to the destination, each node on the establish-
ment path verifies that acceptance of the new channel
is consistent with the guarantees given by the node to
any existing channel. If so, a suggested value of the
local delay bound for this channel is included by the
node in the establishment request. The destination
does the final allocation of the local delay bounds; it
may increase the local delay bounds for the intermedi-
ate nodes but cannot decrease them. These local
delay bounds are assigned to the nodes during the
return trip of the establishment message. Each inter-
mediate node also offers an upper bound on the
amount of buffer space it can reserve for the new
channel. The destination verifies that the amount of
buffer space is indeed sufficient for the channel with
its final delay bounds, and, if possible, reduces the
amount of buffer space required at the intermediate
nodes.
Three tests are made at each intermediate node
during the forward establishment request.
The deterministic test verifies that there is
sufficient bandwidth and processing power at the
node. It is done by verifying that
i
Σti /xmin,i ≤ 1 , (1)
where i ranges over all the channels in the node,
including the new one.
The delay bound test verifies that the delay
bounds assigned to already existing channels can be
satisfied after accepting the new channel. Suppose the
scheduling in the node is deadline based (as described
later in this section). From the perspective of a chan-
nel k, the worst possible arrival pattern on the dif-
ferent channels is one which would cause the deadline
of some packet on channel k to be missed. It is possi-
ble to determine this worst-case situation for each of
the existing channels in the node, and to obtain the
value of a lower bound on the new channel’s delay
bound, so that existing delay bounds are not violated.
For further details, we refer the reader to [Ferrari 89].
The buffer space test verifies that there is
sufficient buffer space in the node for the new chan-
nel. In general, the buffer space required for the new
channel depends on both the local delay bounds and
the traffic characteristics of the channel. Since during
the forward trip, the delay bound is not known, the
node can use an upper bound (for example, the end-
to-end delay) for the purpose of computing the
required buffer space. This allocation can be reduced
when the final bounds are known. For details, we
refer the reader to [Ferrari 90b].
Scheduling
The real-time establishment scheme assumes
that scheduling in the hosts and in the nodes will be
deadline-based (a variant of the earliest due-date
scheduling scheme [Liu 73]). Each real-time packet in
the node is given a deadline, which is the time by
which it is to be serviced. Let di,n be the local delay
bound assigned to channel i in node n. A packet trav-
eling on that channel and arriving at that node at time
to will usually 3 be assigned a node deadline equal to
to+di,n.
The scheduler maintains at least two queues:4
one for real-time packets and the other for all other
types of packets and all local tasks. The first queue
has higher priority, is ordered according to packet
deadlines, and served in order of increasing deadlines.
The second queue can be replaced by multiple
queues, managed by a variety of policies.
Rate Control
At channel establishment time, each intermedi-
ate node checks whether it will be able to accept
packets at the rate declared by the sender. However,
malicious users or faulty behavior by system com-
ponents could cause packets to arrive into the network
at a much higher rate than the declared maximum
value, 1/xmin. This can prevent the satisfaction of
the delay bounds guaranteed to other clients of the
real-time service. A solution to this problem consists
of providing distributed rate control by extending the
deadlines of the ‘‘offending’’ packets. The deadline
assigned to an offending packet would equal the dead-
line that packet would have if it had obeyed the xmin
constraints declared at connection establishment time.
As a consequence of this rate control scheme, an
intermediate node can assume that the clients are
obeying the promised traffic specifications even when
two packets sent at an interval longer than or equal to
xmin by the client come closer together because of
network load fluctuations. This rate control scheme
requires that the nodes downstream allocate sufficient
buffers to provide for any such unintentional viola-
tions of the xmin guarantees.
Let us call a node that implements the rate con-
trol, scheduling, and admission control mechanisms
described above a bounded-delay server. A
bounded-delay server will ensure that no packet along
a channel will spend more than its delay bound in the
node, provided that the channel does not send packets
faster than its specified rate.
¡¡¡¡¡¡¡¡¡¡¡¡¡¡¡¡¡
3 There may be some exceptions due to rate control.
4 There are two types of real-time queues in [Ferrari 89],
one for packets with deterministic or absolute delay bounds,
and the other for packets with probabilistic delay bounds.
Here, as in the rest of the scheme, we have ignored the second
type of traffic to simplify our presentation.

THE JITTER SCHEME
The jitter scheme is based on the establishment
scheme outlined in the previous section. We utilize
the fact that a bound on delay is a bound on delay
variation as well. While the global end-to-end delay
bound (usually a few tens of milliseconds) is too large
to be sufficient as a delay jitter bound, the local delay
bound at one single node (a few milliseconds) can
serve as a good bound on the jitter of a real-time
channel.
In order to provide a delay jitter guarantee, we
need sufficiently faithfully to preserve the original
arrival pattern of the packets on the channel. If this
arrival pattern is faithfully preserved at the last
bounded-delay node on the channel’s path, the max-
imum delay variation of packets on the channel will
equal the delay bound at the last node.
Each node in our jitter control scheme needs to
perform two functions: reconstruct and preserve the
original arrival pattern of packets along a channel,
and ensuring that this pattern is not distorted too
much, so that it is possible for the next node down-
stream to reconstruct the original pattern. Thus, each
node can be looked upon as consisting of a number of
logical components: regulators, one for each of the
channels passing through the node, each responsible
for reconstructing and preserving the arrival pattern of
packets along that channel; and a bounded-delay
server, shared by all channels and ensuring that the
maximum distortion introduced in the arrival pattern
at the next node is bounded. This node model is
shown in Figure 1.
Channel 2
Channel 1
Channel 0
Regulators Bounded-delay server
Figure 1: The node model for jitter control
Bounding the delay jitter according to this
scheme requires the addition of one performance
parameter. Thus, clients must specify the values of the
following two parameters for their performance
requirements:

the source-to-destination delay bound D
for the channel’s packets, and

the source-to-destination delay jitter
bound J for the channel’s packets.
The delay of a packet on channel i should not be less
than Di−Ji, and not greater than Di. If no value of J
is specified, we set as a default J equal to D.
We call a real time channel with guaranteed
delay and jitter bounds (that is, one with different J
and D bounds) a deterministic jitter channel, while
channels with the same value of J and D are simply
called deterministic channels below.
The structure of the regulator module and the
channel establishment scheme are described in the
next sections.
Establishment Of Bounded Jitter Channels
As in the original scheme, the establishment
procedure consists of tests to be performed during the
forward trip of the establishment message with each
of the intermediate nodes proposing some perfor-
mance bounds and the destination node relaxing these
bounds, if possible. The purpose of the establishment
procedure is to determine the local delay bound and
local jitter bound at each of the intermediate nodes.
For channel i, with an end-to-end delay bound Di and
an end-to-end jitter bound Ji, we need to determine
the local delay bound di,n and the local jitter bound
Ji,n at each intermediate node n, which has to ensure
that every packet on channel i has a local delay
greater than di,n − Ji,n but less than di,n.
The paradigm followed is similar to that of the
original scheme: each intermediate node offers a sug-
gested value for the performance bounds on the for-
ward trip, and the destination relaxes these bounds, if
possible. Thus, three values need to be proposed for a
channel being established on the forward trip, a sug-
gested delay bound, a suggested jitter bound, and a
suggested bound on the buffer space available at the
node. We, however, require that the node always
offer the same values for the local delay bound and
the local jitter bound. Thus, only two bounds, the
bound on buffer space and the bound on delay, need
to be suggested by each intermediate node during the
forward trip. These are the same bounds as in the ori-
ginal scheme, but the buffer bound computation is
done in a different fashion (described below). The
delay bound is computed as in the original scheme,
and is interpreted as a lower bound on the delay jitter
offered by the node as well.
The regulator corresponding to this channel in
the node immediately downstream (see Figure 1) is
responsible for restoring the original arrival pattern
that the channel’s traffic had when it entered the net-
work by absorbing the jitter introduced by this node.
By reconstructing the original arrival pattern at each
intermediate node, the jitter on the channel can be
controlled.
Thus, the following two tests will need to be
performed at each of the intermediate nodes:
Jitter test: this would comprise the deterministic
test and the delay bound computations (except of
course that they are now jitter bound computations).
The latter would return the value of the least possible

jitter that the node can provide to the new channel.
Buffer space test: this determines whether there
is sufficient space to accommodate the new channel,
and how much of the existing space should be
reserved for it.
The buffer space required to prevent any losses
consists of two factors: (a) the amount of buffer space
required because of the local delay bound at the node,
which dictates how long a packet may stay in the
node; and (b) the amount of buffer space required to
absorb the jitter in the previous node and to recon-
struct the original arrival pattern. Assuming the
correctness of the algorithms and tests in the original
scheme, no packet (after the original arrival pattern is
reconstructed) will stay in the node for more than the
delay bound in that node. Since the reconstruction is
done by delaying each packet so as to absorb the
delay jitter from the node immediately upstream, no
packet will be held longer than the local delay jitter
bound at the previous node in this reconstruction pro-
cess. It follows that the number of buffers bi,n needed
to ensure that no packets from channel i will be lost
at node n is
bi,n = smax,i

di,n /xmin,i

+
!
Ji,n −1 /xmin,i
#$%
,(2)
where di,n is the delay jitter bound assigned to chan-
nel i at node n and Ji,n −1 is the jitter bound assigned
to channel i at the previous node.
The buffer space test consists of determining if
there is sufficient space (as given by equation 2)
available in the node for the new channel. However,
at the time of the forward trip, neither the final value
of di,n, nor that of Ji,n −1 are known. Thus the node
must assume an upper bound for these values. The
simplest way is to bound the sum of both these values
by the end-to-end delay requirement Di of channel i.
If the buffer space required by this (admittedly crude)
guess is not available, the amount of space available
(or a fraction thereof) is temporarily assigned to the
channel.
The Destination Host Tests And Algorithms
The destination host, in the modified scheme,
has to perform additional tests in order to assure that
the end-to-end channel jitter and delay bounds are
met. The jitter bound offered by the last node on the
path (say the Nth node) must satisfy
Ji,N ≤ Ji. (3)
The destination must determine whether the
delay requirement of the new channel can be met by
the nodes along the path. It can do so by verifying
that, for channel i with total delay bound Di,
Di ≥
n =1
Σ
N
Ji,n
l
. (4)
This is a variation of the D test of [Ferrari 90a].
The destination now has the responsibility of dividing
the delay bound and the jitter bound among the inter-
mediate nodes. A very simple way to do this (distri-
buting delays in the same manner as in the original
scheme) would be
di,n =
N
1'
()
0 Di −
m =1
Σ
N
Ji,m
l
12
3 + Ji,n
l
, (5)
Ji,N=Ji, (6)
Ji,n=di,n, (7)
where n ranges over all the intermediate nodes on the
channel’s path, except the last (Nth) node.
The final test consists of verifying that the
buffer space allocated to the channel by the inter-
mediate nodes along the path is sufficient to ensure
that packets will not be lost. This consists of recom-
puting the buffer space requirement at each node
according to equation (2), and verifying that the
recomputed amount is available in the node.5
Rate Control And Scheduling
While the scheduling policy is unchanged in the
bounded-delay node(see Figure 1), a new rate control
algorithm is performed at each of the newly intro-
duced regulators. The new rate control mechanism is
used to restore the arrival pattern of packets that is
distorted in the previous node by network load
fluctuations. The bounded-delay server needs to per-
form one more task than in the old scheme: it has to
write in each packet’s header the difference between
the instant the packet is served, and the instant it was
supposed to be serviced (its deadline). This informa-
tion will be used by the regulator immediately down-
stream.
Let di,n be the local delay bound assigned to
channel i in node n, and Ji,n be the local jitter bound
assigned to the channel in that node. A packet travel-
ing on channel i that is subjected to the maximum
possible delay at node n−1 and arriving at node n at
time to will usually be assigned a node deadline equal
to to+di,n, and a node eligibility-time equal to
to+di,n−Ji,n. The packet is ineligible for transmission
until its eligibility-time, which ensures that the
minimum delay requirements for channel i are met in
node n. Usually the value of di,n will be the same as
Ji,n; hence, the eligibility-time will be the same as the
arrival time at an intermediate node, and a packet will
be eligible for transmission as soon as it arrives. Any
packet arriving closer than the specified value of xmin
4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4¡4
5 It is possible to play with the values of delay and buffer
space bounds and to devise more sophisticated allocation
schemes that would minimize the possibility of rejection by
the destination host tests, but we will not discuss them in this
paper.

to the previous packet on the same channel is made
ineligible for a longer period of time, up to the time it
was supposed to arrive obeying the promised
minimum inter-arrival time xmin. Moreover, the
difference between the actual time that the packet was
serviced in the previous node and the deadline in the
previous node is read from the packet’s header (which
was stamped there by the previous node), and the
packet’s eligibility-time (i.e., the time when the
packet will be put into the scheduler queue) is
increased by this amount. In effect, this extension of
the eligibility-time forces each packet on the channel
to behave as if it experienced a constant amount of
delay, the bound of di,n −1 at the previous node. The
difference between a packet’s eligibility-time and its
deadline always remains the same as the channel’s
jitter bound at the node.
Let the holding time of a packet be defined as
the period that the packet is ineligible for service after
its arrival. The pseudo-code implementing the above
rate control scheme is shown in Figure 2.
Real-time packets which are ineligible for
transmission are kept in a queue from which they are
transferred to the scheduler as they become eligible.
This queue can be maintained as a set of calendar
queues [Brown 88] which can be made very fast by
hardware implementation; packets are inserted in a
queue indexed by their eligibility-time, and all the
packets that are in the queue indexed by the current
time become eligible.
One important consequence of this rate control
scheme is that the arrival pattern of real-time packets
entering the scheduler at any intermediate node is
identical to the arrival pattern of these packets at the
entry point to the network, provided the client obeyed
the xmin constraint. As a result, the deadline assigned
to a packet in a node is given by the time it entered
the network plus a constant amount, the constant
being the sum of the delay bounds assigned to the
nodes along the partial route of the real-time channel
covered so far. If di,k is the delay bound assigned to
channel i at the kth node along its path, the deadline
dln assigned in node n to a channel i packet which
entered the network at time to is
dln=to +
k=1
Σ
n
di,k + Pn, (9)
where Pn is the propagation delay from the source till
node n.
As a result of equation (9), the jitter of packets
on a real-time channel at its exit point from the net-
work equals the jitter introduced by the last node in
the network, which leads to the justification of the test
in equation (3).
565757575757575757575757575757575757575757575757575757575757575757575757575757575757575
correction term 898898 deadline in previous node
- actual completion time
in previous node
holding time @9@@9@ correction term
+ delay bound
- jitter bound at this node
eligibility-time A9AA9A holding time
+ arrival time
deadline B9BB9B max[{eligibility-time+jitter bound},
{deadline of last packet+xmin}]C6C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C7C
Figure 2: The rate control algorithmD6D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D7D
Correctness
Before attempting a formal proof of correct-
ness, let us try to give an intuitive explanation why
our jitter scheme works. The goal in our scheme is to
have a constant delay-line between the sender and the
receiver. However, the queueing nodes introduce
jitter because the queueing delay in the bounded-
delay server is variable. The jitter introduced by each
bounded-delay server is absorbed by the regulator at
the next node (see Figure 1), and thus, the bounded-
delay server and the regulator together provide a
constant-delay element. A real-time channel passes
through a number of these constant-delay elements,
and thus the delay jitter seen by the receiver is merely
the delay jitter introduced by the last bounded-delay
server along the path.
In order to show that the scheme is correct, and
that packet jitter is indeed bounded, we prove two
lemmas and then assert a theorem concerning correct-
ness.
Lemma 1: Equation (9) is valid for the first
packet of a channel at any node along the path of an
established channel. Proof: For the first packet,
which entered the network at time t1, the result holds
trivially at the first node. As this packet moves on to
the second node, it carries with it the correction term,
which is the difference between the time it was ser-
viced and its deadline. Suppose the packet was ser-
viced at time instant t′; then the correction term is
t1+d1−t′. Assuming that the constant propagation
delay has been filtered out, the packet reaches the
second node at time t′ and is assigned a deadline of
t′+d2+t1+d1−t′ according to Figure 2. Thus, the
deadline assigned is t1+d1+d2. It is easy to show
that the same result holds for other nodes along the
path by induction on the path length. The propagation
term also builds up as we proceed along the path of
the channel.6
Lemma 2: If a channel client does not send
packets faster than the promised interval xmin, equa-
E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E¡E
6 If the propagation delay is not a constant on all links, the
jitter introduced by the links must also be accounted for. One
way to do so would be to add this delay jitter to the right-hand
side of equation (6).

tion (9) is valid for all packets on the channel at all
nodes along its path. Proof: We already know that
equation (9) is valid for the first packet on the channel
at any node along the path. Let us consider the second
packet along the channel, which enters the network at
time t2. Since the client is well-behaved, the packets
never enter the network at intervals shorter than xmin.
Recalling the fact that the rate control algorithm of
Figure 2 would execute the same set of instructions as
for the first packet, the deadline assigned in the last
step would equal that given by equation (9), unless it
were modified by the comparison to the previous
packet’s deadline in the last line of Figure 2. Since
the client is well-behaved, the difference between t2
and t1 would be at least xmin, hence equation (9)
applies. Similarly at all the other nodes along the path.
Thus, equation (9) is shown to be valid for the second
packet and by induction can be extended for all the
packets on the channel.
Having proved Lemma 2, we can now state
Theorem 1: If a packet on an established real-time
channel enters the network at time t and the client
does not violate his traffic specifications, the packet is
delivered by the network to the receiver in between
the time instant t+D−J and the time t+D.
Proof: From Lemma 2, the deadline assigned at
the last node to a packet that enters the network at
time instant t must be t+D, since the local delay
bounds at all the intermediate nodes sum up to the
end-to-end delay bound. (It can be verified from equa-
tions (5) and (6) that it is indeed the case.) Thus, the
deadline assigned at the last node is t+D, and the
jitter bound assigned by the jitter control algorithm at
the last node according to equation (7) is J. Since the
client is well-behaved and did not violate his traffic
specifications, the last step of the rate control algo-
rithm would assign a deadline equal to the sum of the
eligibility-time and the jitter bound. It follows then
that the eligibility-time of the packet in the last node
was t+D−J. Since the packet is handed to the
bounded-delay server (the scheduler) only at the
eligibility-time and the bounded-delay server always
meets the deadlines of the packets in the queue, it fol-
lows that the packet can leave the last node only in
between the time instants t+D−J and t+D.
In the next section, we offer a quantitative
measure of the effectiveness of the scheme obtained
by means of simulations.
THE SIMULATIONS
In the previous sections, we have presented our
establishment scheme together with a new distributed
rate control mechanism. In this section, we will give
simulation results for our scheme and compare the
delay characteristics and buffer requirements for
real-time channels with and without jitter control.
Our goal is to provide simulation-based
answers to the following questions:
S 0 1 2 3 4 5 R
Cross Traffic
Figure 3: The simulated network. S is the
sender and R is the receiver for the real-time
channels under observation.
F
What is the delay distribution of packets on
deterministic jitter channels? How does it com-
pare to that on deterministic channels without
jitter control?
G
What are the buffer requirements and the effec-
tive utilization of the allocated buffer space
with the new rate control mechanism? How do
they compare with those obtained using the old
rate control mechanism?
H
What are the delay characteristics of deter-
ministic channels with the new rate control
mechanism but without jitter bounds?
Our simulations were based on a simulator writ-
ten using the simulation package CSIM [Schwetman
89].
We have simulated networks with several dif-
ferent topologies and traffic configurations. In all
cases in which there was an appreciable delay jitter,
our scheme was found capable of reducing it. We
shall present the results we obtained for the network
shown in Figure 3, where uncontrolled jitter may be
substantial.
The jitter control scheme proposed in this paper
differs from the previous work [Ferrari 90a] in two
aspects: a new rate control mechanism has been pro-
posed, and establishment tests have been revised to
account for the jitter requirement explicitly: the revi-
sion consisting of jitter allocation according to equa-
tions (6) and (7).
To study the effects of both changes, we exam-
ined the behavior of three channels that traverse the
6-node path shown in Figure 3: a channel using the
old rate control scheme(channel A), a channel using
the new rate control, but with no jitter
allocation(channel B), and a channel with both the
new rate control and the assignment of jitter bounds
according to equations (6) and (7). The three channels

had identical characteristics; both the service time (t)
and the packet size (smax) were taken to be one unit
each, xmin was 20 units, and the delay bound was 144
units. The jitter requirement for channel C was 7 time
units. Since there were 6 nodes along the path, the
average local delay bound for one channel in each
node was about 24 time units.
In each node, there were also about 20 cross
channels with local delay bounds ranging from 5 to 25
units, xmin, t and smax being the same as those for
channels A, B and C. The traffic on all the channels
(the three channels under study and cross channels)
was generated using an on-off model, in which pack-
ets were only generated in ‘‘on’’ mode and the ratio
between the time spent in the n’’ mode to that in the
‘‘off’’ mode ranged from 6 to 10 for different chan-
nels. The addition of cross traffic caused the nodes to
have a total utilization of about 0.8.
Histogram of end-to-end delays
0 20 40 60 80 100 120
0
0.02
0.04
0.06
0.08
0.1
.94
.44
delay
bound
A B C
Figure 4: The delay distribution of packets with
and without jitter control. Channel A using the
old rate control has a substantial amount of de-
lay jitter, which is alleviated somewhat by us-
ing the new rate control scheme for channel B.
The jitter is further reduced by jitter allocation
using equations (6) and (7) for channel C. The
numbers 0.44 and 0.94 in the figure indicate
frequency counts that are too large to be shown
on the graph.
Figure 4 shows the delay distributions of pack-
ets on the three channels. Channel A has a much
larger spread of delays, but these delays are much
lower than those of channel B or C. Thus, the receiv-
ing node will have to buffer these packets for a much
longer period to delay the packets that have arrived
too early, and this can lead to excess buffer require-
ments. A large number of buffers have to be provided,
even though a very small number of packets experi-
ence the large delays that require the reservation of
buffer space in the destination. Clearly, shifting the
bounds as close as possible to the experienced delays
looks a desirable property from this viewpoint. The
new rate control and jitter allocation for channels B
and C reduce the spread at the cost of introducing
extra delays in the network.
Before presenting the simulation results for the
allocated buffer space, we will perform a simple
analysis. Assume channel i passes through nodes
1,... N, the source node being 1 and the destination
node being N. In the worst case, there could be
n =1
Σ
k
di,n /xmin,i packets from channel i at node k, so
this much buffer space needs to be allocated at node
k. The reader will observe that an increasing amount
of buffer space is needed as we proceed along the
path of the channel. For a channel with jitter control
using the modified rate control mechanism, however,
the maximum time a packet can stay at node k is
Ji,k−1+di,k, where Ji,k−1 is the maximum jitter in
node k −1 and di,k is the maximum delay bound in
node k.
Figure 5 shows the allocated buffer space and
the actual buffer utilization for real-time channels
with and without jitter control. The observed results
roughly match the above analysis, the differences
being due to the fact that network dynamics can cause
different delay and jitter bounds to be assigned along
the path of a channel. In channels without jitter con-
trol, the buffer space allocated in each node increases
as we move away from the source. As can be seen,
the buffer requirements at a node using the proposed
rate control mechanism with jitter control does not
increase with the path length, unlike what happens to
channels without these jitter control mechanisms.
Our scheme results in a savings of buffer space when-
ever the path consists of roughly three or more nodes.
For jitter-controlled channels, the allocated buffer
space is almost constant at all nodes. It is shown in
Figure 5 that the actual number of buffers used by
both the channels A and C is around 2 in the inter-
mediate nodes, but channel A uses many more buffers
at the receiver. Figure 5 also shows that almost all the
buffer space allocated in the destination node was
actually needed in our simulations.
CONCLUSIONS
This study was undertaken to determine the
feasibility of offering delay jitter guarantees in
packet-switching networks. We defined the problem
and the type of network to be studied, adopted the
real-time channel abstraction capable of providing
bounded delay and loss rate guarantees, and extended
it to provide delay jitter guarantees, using a traffic
characterization already used in previous studies in
this area. Our delay jitter control algorithm includes a
new rate control mechanism and the corresponding
modified establishment tests. Unlike pure circuit

Buffers
Relative position in path
0 1 2 3 4
0
2
4
6
8
10
R
I . . . . . . .P . . . . . . .Q. . . . . . .R . . . . . . .S ................T
∗. . . . . . .∗. . . . . . .∗. . . . . . .∗. . . . . . .∗. . . . . . .∗
. . . . . . .
. . . . . . .∗
U
Allocation for A
Utilization by A
Allocation for C
Utilization by C
Figure 5: Buffer requirements and utilization
with and without jitter control. This figure illus-
trates buffer allocation and utilization for a
deterministic channel (without jitter control)
and a deterministic jitter channel. The deter-
ministic channel A has 5 hops, and as can be
seen, the allocated buffer space in a node in-
creases linearly with its position along the path.
The deterministic jitter channel C also has 5
hops, but the allocated buffer space for each
host on the path is almost constant. The buffer
space allocated in intermediate nodes for the
deterministic channel is almost unused, while
the buffer utilization for the deterministic jitter
channel is, again, constant. If we want to con-
trol delay variance for a deterministic channel
by buffering packets in the end, we get a
significant amount of buffer utilization in the
last node. R is the destination node.
switching, we are able to provide different jitter
guarantees to different connections, and utilize the
network resources to a fair extent.
Our scheme has been proven to be correct, and
simulations have shown that it can be very effective
in reducing delay jitter. One advantage of the variance
control schemes we have presented is that the amount
of buffer space required for real-time channels to
prevent packet losses is significantly reduced.
Another significant feature is that the jitter provided
by our scheme is independent of the length of a
channel’s path; this allows us to offer bounds on delay
jitter significantly smaller than the global delay
bounds of a real-time channel.
Some of the work that remains to be done in the
area of delay jitter control consists of :
V
a scheme that provides delay jitter guarantees
while accounting for the burstiness of traffic;
W
a scheme to provide for fast establishment of
real-time channels with delay jitter control;
X
a comparison of the cost of providing delay
jitter guarantees by our scheme in a packet-
switched network to the cost of a circuit-
switched network; and
Y
a scheme that can offer delay jitter guarantees
without offering a delay guarantee.
REFERENCES
[Brown 88] R. Brown, Calendar Queues: A Fast
O(1) Priority Queue Implementation for the Simula-
tion Event Set Problem, Communications of the
ACM, vol. 31, n. 10, pp. 1220-1227, October 1988.
[Ferrari 89] D. Ferrari, Real-Time Communication in
Packet Switching Wide-Area Networks, Tech. Rept.
TR-89-022, International Computer Science Institute,
Berkeley, May 1989, submitted for publication.
[Ferrari 90] D. Ferrari, Client Requirements for
Real-Time Communication, IEEE Communications
Magazine, vol 28, no 11, pp. 65-72, November 1990.
[Ferrari 90a] D. Ferrari and D. C. Verma, A Scheme
for Real-Time Channel Establishment in Wide-Area
Networks, IEEE J. Selected Areas Communic., vol.
SAC-8, n. 3, pp. 368-379, April 1990.
[Ferrari 90b] D. Ferrari and D. C. Verma, Buffer
Allocation for Real-Time Channels in a Packet-
Switching Network, Tech. Rept. TR-90-022, Interna-
tional Computer Science Institute, Berkeley, June
1990, submitted for publication.
[Leiner 89] B. Leiner, Critical Issues in High
Bandwidth Networking, Internet RFC-1077, Nov.
1988.
[Liu 73] C. L. Liu and J. W. Leyland, Scheduling
Algorithms for Multiprogramming in a Hard Real
Time Environment, J. ACM, vol. 20, n. 1, pp. 46-61,
January 1973.
[Schwetman 89] H. Schwetman, CSIM Reference
Manual (Revision 12), MCC Tech. Rept. No. ACA-
ST-252-87, January 1989.

Delay jitter control for real time communication

More Related Content

What's hot (19)

Viewers also liked (9)

Similar to Delay jitter control for real time communication (20)

Recently uploaded (20)

Delay jitter control for real time communication