2003 scalable networking - unknown

Scalable Network Programming
Or: The Quest For A Good Web Server (That Survives Slashdot)

Felix von Leitner
felix-linuxkongress@fefe.de

2003-10-16

Abstract
How do you write a fast network server?
How do you write a network server that can handle 10000 clients?
What bottlenecks are there and how do you avoid them?



About me

I do security consulting for a living (my company is called Code Blau).

I am particularly interested in small and high performance code.

Besides writing web servers, I also wrote an ftp server, a very fast LDAP
server (read-only so far, I’m waiting for someone to pay me to complete it), and
the smallest Unix libc (which I also used to compile and run all the programs
mentioned in this talk).

I’m also known in some circles for porting djbdns to IPv6, writing the FAQ
for the mutt mail client, and for writing inﬂammatory bug reports on the Linux
kernel mailing list.

Scalable Network Programming 1


Why care about high performance network code?

Apache ist fast enough for most things.

But if you read Slashdot a lot, you will see that the sites they link to are
often down, because they can’t handle the load.

Slashdot itself uses 8 P3/600 with 1 Gig RAM and 10k rpm SCSI disks.

www.heise.de uses 4 P3/650 with 1 Gig RAM.

ftp.fu-berlin.de is an SGI Origin 200 with 2 R10k/225 with 1 Gig RAM and
10k rpm SCSI disks.

It’s obviously possible to buy performance with more hardware, but how
can we make sure the software is not the bottleneck?



Why is it important to handle many connections?

I was called by an internet toy shop once. It was a rainy day in December.

They are making most of their money in the Christmas season.

They called because their web server was under attack. Some sort of
distributed attack had opened 100.000 HTTP connections to their web server.

This caused 20.000 Apache processes to be spawned on each of their load
balanced Solaris back-ends. The machines were desparation swapping.

It was impossible to buy anything from them.

I think they were actually afraid of bankruptcy.



First, let’s write a web client

char buf[4096];
int len;
int fd=socket(PF_INET,SOCK_STREAM,IPPROTO_TCP);
struct sockaddr_in si;

si.sin_family=PF_INET;
inet_aton("127.0.0.1",&si.sin_addr);
si.sin_port=htons(80);
connect(fd,(struct sockaddr*)si,sizeof si);
write(fd,"GET / HTTP/1.0rnrn");
len=read(fd,buf,sizeof buf);
close(fd);



That’s it?

Well,

1. I didn’t #include all the headers

2. there is no error handling

3. the client can only fetch up to 4k (including HTTP headers)

4. you can’t actually specify the URL

but other than that: yeah, that’s pretty much it.



OK, then let’s write a web server!

int cfd,fd=socket(PF_INET,SOCK_STREAM,IPPROTO_TCP);
struct sockaddr_in si;

si.sin_family=PF_INET;
inet_aton("127.0.0.1",&si.sin_addr);
si.sin_port=htons(80);
bind(fd,(struct sockaddr*)si,sizeof si);
listen(fd);
while ((cfd=accept(fd,(struct sockaddr*)si,sizeof si)) != -1) {
read_request(cfd); /* read(cfd,...) until "rnrn" */
write(cfd,"200 OK HTTP/1.0rnrn"
"That’s it. You’re welcome.",19+27);
close(cfd);
}



This server sucks!!!

This server (besides not actually implementing the protocol) has the
drawback that it can only handle one client at a time. Here is a better one:

while ((cfd=accept(fd,(struct sockaddr*)si,sizeof si)) != -1) {
if (fork()>0) continue; /* handle connection in a child process */
read_request(cfd); /* read(cfd,...) until "rnrn" */
write(cfd,"200 OK HTTP/1.0rnrn"
"That’s it. You’re welcome.",19+27);
close(cfd);
exit(0);
}



One process per connection – is that a good idea?

That’s the old school way. It’s been done this way since the 70ies.
The ﬁrst web server from CERN used this method.



One process per connection – is that a good idea?

Creating one process per connection does have scalability issues.

It is difﬁcult to implement fork() properly. Here is a benchmark:

pipe(pfd);
for (i=0; i<4000; ++i) {
gettimeofday(&a,0);
if (fork()>0) {
write(pfd[1],"+",1); block(); exit(0);
}
read(pfd[0],buf,1);
gettimeofday(&b,0);
printf("%llun",difference(&a,&b));
}



latency: forking a process
3000
Linux 2.4
NetBSD 1.6.1
Linux 2.6
FreeBSD 5.1
2500 OpenBSD 3.4

2000
microseconds

1500

1000

500

0
0 500 1000 1500 2000 2500 3000 3500 4000
number of processes



fork: dynamic linking versus static linking

Another important fork performance factor is how much fork has to do.

On modern platforms with a hardware memory management unit, fork only
copies the page mappings.

Dynamic linking creates many page mappings for the ELF sections in the
shared libraries, and the Global Offset Table and so on.

So, linking the fork benchmark statically ought to improve performance
noticeably.



latency: forking a process
1000
Linux 2.6 static
Linux 2.6 dynamic
900 FreeBSD 5.1 static
FreeBSD 5.1 dynamic

800

700

600
microseconds

500

400

300

200

100

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
number of processes



Just to make sure you understood these numbers

The fork-and-do-something latency on my notebook on Linux 2.6 is 200
microseconds.

That means my notebook can create 5.000 processes per second.

Thus my notebook can handle about 13 billion forks per month.

My Athlon XP 2000+ desktop can do 10.000 processes per second, or 26
billion per month.

Heise Online, the biggest German site, had 118 million page impressions in
September.



Scheduling

Why does the fork benchmark include writing to the pipe?

Because having many processes not only makes creating more processes
more difﬁcult, it also makes choosing which process to run more difﬁcult.

The part of the operating system that chooses which process to run next is
called the scheduler.

A typical workload is having two dozen processes, with one or two of them
actually having something to do (they are runnable).



Scheduling

Linux interrupts running processes every 1/100th second (1/1000th on
Alpha; this value is traditionally called HZ on Unix and Linux and is a compile
time constant) to give others a chance to run.

The scheduler is also activated when a process blocks (waiting for input, for
example).

The scheduler’s job is to select the process that should run next. What
makes this hard is the fairness requirement: all processes should get an equal
share of the CPU. In particular, no process should starve.

It obviously makes sense to have two lists: one for the runnable processes,
one for the blocked ones.



Scheduling

Unix has a mechanism to favour interactive processes over batch jobs. To
that end, the kernel calculates a nice value for each process once per second.

When there are 10.000 processes, this thrashes the 2nd level cache.

This is especially bad for SMP, because the process table must be protected
by a spinlock, i.e. the other CPUs can’t switch processes while the ﬁrst CPU
calculates the nice value.

The typical solution from the commercial Unices is to have one run queue
per CPU. Further optimizations include keeping the run queues sorted, for
example with a heap, or having one run queue per priority.



The Linux 2.4 Scheduler

The Linux scheduler has one unsorted run queue with all runnable
processes and one task list for all sleeping processes and zombies.

All operations on the run queue are protected by a spin lock.

There are several new experimental schedulers for Linux 2.4, with heap
based priority queues or multiple run queues, but the most amazing scheduler
is the O(1) scheduler by Ingo Molnar.

The O(1) scheduler is the default scheduler in Linux 2.6.



The Linux 2.6 Scheduler

The O(1) scheduler has two arrays of pointers to a linked list for each CPU.
The arrays have one entry for each possible priority.

The linked lists contain all the tasks for the given priority. Since all the
tasks in each list have the same priority, putting a process of a given priority
into this data structure takes constant time.

One of the arrays is the current run queue; all the processes in it are run in
turn, and when they used up their time slice they are moved to the other array.
When the run queue is empty, the arrays are swapped.

Processes that the scheduler interrupted are assumed to be batch processes
and punished. That is cheaper than favouring interactive processes.



How important is scheduler performance?

The scheduler is run 100 (HZ, actually) times per second, plus each time a
process is blocking.

Each time the scheduler is suspending one process and switching to
another one, the operating system increases a counter.

This ”context switch” counter can be seen using vmstat.

It is architecture dependent how expensive a context switch is. Basically,
the more registers, the more expensive the context switch. So x86 actually has
an unfair advantage over RISC, in particular over SPARC and IA64.



How important is scheduler performance?

Here is a vmstat, running an MP3 player:

procs memory swap io system cpu
r b w free buff cache si so bi bo in cs us sy id
0 0 0 60316 9136 316280 0 0 135 113 47 8 8 9 82
0 0 0 60296 9152 316280 0 0 0 0 200 54 3 0 97

And now with httpbench running over lo and eth0:

4 0 0 56136 9236 317936 0 0 24 0 228 7953 6 59 35
0 0 0 56308 9236 317936 0 0 0 0 221 1590 4 10 86

4 0 0 53476 9304 320180 0 0 0 0 16416 12319 26 74 0
5 0 0 52552 9304 320180 0 0 0 0 16391 12307 20 80 0



Other problems with running many processes

Creating very many processes has another problem: memory consumption.

Each process takes up memory. When a process forks a child, all the pages
are set to copy-on-write in both processes. As soon as a process writes to a
page, it gets copied.

This is good, but most people don’t realize just how much their programs
write all over the main memory.

For example, calling malloc or free may cause a chain reaction of coalescing
data structures or rebalancing some tree in the free list.



Other problems with running many processes

Another example is the dynamic linker, which maintains the Global Offset
Table, but does so lazily per default, i.e. the offsets in the table are updated
only when the function is ﬁrst called.

So if you call a function the ﬁrst time after a fork, and you fork 1000
children, then you will have wasted 1000 4k pages.

You can tell the dynamic loader to update all offsets at program start by
setting the $LD_BIND_NOW environment variable.

Please note that cleaning up 1000 processes is usually slower than forking
1000 processes, and all the tested systems are unresponsive during that time.



Memory Consumption

Measuring your own memory consumption is not so easy on Unix.
There is a getrusage system call to do it, but Linux always returns 0
for the memory usage, and FreeBSD only returns data for dynamically linked
programs.
On Linux, /proc/self/statm contains the data:

% ../show/memusage-glibc-c
1164K total size, 324K resident, 1132K shared
% ../show/memusage-glibc-static-c
% ../show/memusage-diet-c



The one-process-per-connection model

This programming model works so well, there are standard tools for it.

Unix comes with the inetd server that not only forks a new process for
each connection, it also executes an external program for each connection.

Unfortunately, inetd has certain shortcomings that have given it (and the
whole network programming model) a bad name, so some people wrote their
own replacements, for example tcpserver, xinetd, ipsvd.

This is slightly less efﬁcient than simply forking, but it has the advantage
that you can standardize the IP access control mechanism.

I wrote my ﬁrst web server, fnord, using this model. It still serves
www.fefe.de.



HTTP connect latency
2000
Fnord 1.9
Gatling 0.3
1800

1600

1400

1200
microseconds

1000

800

600

400

200

0
0 500 1000 1500 2000 2500
number of open connections



HTTP serve latency (Linux 2.6)
16000
Fnord 1.9
Gatling 0.3
Apache 2.0.47/highperformance.conf
14000

12000

10000
microseconds

8000

6000

4000

2000

0
0 500 1000 1500 2000 2500
open connections



The one-process-per-connection model

Apache also uses the one process per connection model.

However, to save the fork latency, Apache pre-forks, i.e. it forks ﬁrst and
then accepts and delegates the HTTP connection to the child.

But wait! You haven’t seen the whole Apache graph yet!

I clipped away most of it, because then the gatling and fnord measurements
wouldn’t have been visible on the combined graph.

In conclusion: Apache performance sucks.



The complete Apache graph
450000
Apache 2.0.45

400000

350000

300000
HTTP service latency

250000

200000

150000

100000

50000

0
0 200 400 600 800 1000 1200
open connections



Since processes are so slow, why not use threads instead?

It is common knowledge that fork is slow and we should use threads
instead.
The truth is more complex than that.
fork is not universally slow, as you have seen. But on some systems, fork is
really slow. One of these systems is Solaris.
Apparently it was too difﬁcult to ﬁx fork (or all the engineers were busy
inventing Java), so threads were invented instead.
Two years ago, I measured that pthread_create on Solaris is slower than
fork on Linux on the same hardware. However, I lost access to that hardware,
so I don’t have that graph with me today. I do have this:



HTTP request latency; glibc 2.3.2 on Linux 2.6
2500
fork
pthread_create

2000

1500
microseconds

1000

500

0
0 200 400 600 800 1000 1200
open connections



Multithreading performance

One problem with threads is that on a bad OS, creating threads is still slow.
Solaris and Windows may create threads faster than processes, but it is still way
too slow.

So what did they do? They invented thread pools.

A thread pool works like Apache’s pre-forking. The threads are created en
block at the program start, and then the work load is distributed evenly.

When there are more connections than threads, the new connections have
to wait. But on the plus side, you save the cost for the thread creation.

I don’t know about you, but I wouldn’t call that ”scalable”.



Positive aspects of threading

There are some good aspects to Java and threading as well! We got new,
faster hardware, and the RAM prices are on an all-time low!

Same for software: the cretins on the application layer forced great
innovations on the OS layer. Because Lotus Notes keeps one TCP connection
per client open, IBM contributed major optimizations for the ”one process,
100.000 open connections” case to Linux.

And the O(1) scheduler was originally created to score well on some
irrelevant Java benchmark.

The bottom line is that this bloat beneﬁts all of us. We just need to make
sure that there are always small and efﬁcient alternatives to all the crapware.



How is timeout handling done?

One important question for network servers is: how do you detect
timeouts?

A network server needs to detect when a client connects and then does
nothing with the connection.

No matter how well optimized your network server is, you (and the OS)
will always have to keep state for established connections, and those system
ressources are better spent on handling active connections.



How is timeout handling done?

Unix has a standard function called alarm that is well suited for this:

alarm(23); /* deliver SIGALRM in 23 seconds */

Receiving an unhandled signal terminates the program.

So, in 23 seconds, the kernel will terminate the web server.

Later calls to alarm override the previous alarm.

You simply call alarm every time you get activity on the socket.



Timeout handling with select()

select waits for events on one or more ﬁle descriptors (select appeared
1983 in 4.2BSD):
fd_set rfd;
struct timeval tv;
FD_ZERO(&rfd); FD_SET(0,&rfd); /* fd 0 */
tv.tv_sec=5; tv.tv_usec=0; /* 5 seconds */
if (select(1,&rfd,0,0,&tv)==0) /* read, write, error, timeout */
handle_timeout();
if (FD_ISSET(0, &rfd))
can_read_on_fd_0();

The timeout value is given in microsecond resolution, but no Unix actually
offers this accuracy. The ﬁrst argument is the highest fd plus one. *Puke*



Disadvantages of select()

select does not say how long it waited for the events. So you still have to
call gettimeofday() manually. Actually, select does say on Linux, but that is
not portable and actually breaks portable software.

select works on bit vectors. The size of these is OS dependent. If you are
lucky, you get 1024 (on Linux, for example). Many other Unices have less.

This is worse than it looks. Some bad DNS libraries use select for their
timeout handling. If you open 1024 ﬁles, DNS suddenly stops working, because
the DNS socket is over 1024! Apache works around this by manually keeping
the descriptors below 15 free (with dup2).



Timeout handling with poll()

poll is a variant on select (poll appeared 1986 in System V R3):
struct pollfd pfd[2];
pfd[0].fd=0; pfd[0].events=POLLIN;
pfd[1].fd=1; pfd[1].events=POLLOUT|POLLERR;
if (poll(pfd,2,1000)==0) /* 2 records, 1000 milliseconds timeout */
handle_timeout();
if (pfd[0].revents&POLLIN) can_read_on_fd_0();
if (pfd[1].revents&POLLOUT) can_write_on_fd_1();

Pro: no limit on the number of records and descriptors.

Con: poll is not available on some museum exhibitions. Only one current
”Unix” disqualiﬁes itself for networking by not offering poll: MacOS X.



Disadvantes of poll()

The whole array is unnecessarily copied around between user and kernel
space. The kernel then ﬁnds out about the events and sets the corresponding
revents.

Current CPUs spend most of their time waiting for memory anyway. poll
makes that even worse, and the wasted effort rises linearly with the number of
ﬁle descriptors.

This is not as bad as it looks. If poll takes longer, the events are not lost.
The kernel will queue the events and signal them on the next poll.

On the other hand, on Linux and FreeBSD a fork based web server actually
scales better than a poll based one.



Linux 2.4: SIGIO

Linux 2.4 can tell you about poll events via signals as well.

int sigio_add_fd(int fd) {
static const int signum=SIGRTMIN+1;
static pid_t mypid=0;
if (!mypid) mypid=getpid();
fcntl(fd,F_SETOWN,mypid);
fcntl(fd,F_SETSIG,signum);
fcntl(fd,F_SETFL,fcntl(fd,F_GETFL)|O_NONBLOCK|O_ASYNC);
}

int sigio_rm_fd(struct sigio* s,int fd) {
fcntl(fd,F_SETFL,fcntl(fd,F_GETFL)&(~O_ASYNC));
}



Linux 2.4: SIGIO

SIGIO is no direct poll replacement. poll tells you when a descriptor is
readable. SIGIO tells you when it becomes readable.

If poll says you can read from ﬁle descriptor 3 and you don’t do anything
about it, the next call to poll will tell you again. SIGIO won’t. The poll way is
called level triggered, the SIGIO way is called edge triggered.

The best way to get the events is sigtimedwait, after you block SIGIO
(the signal) and signum. That restores synchronicity and avoids the need for
locking and reentrant functions.



Linux 2.4: SIGIO

for (;;) {
timeout.tv_sec=0;
timeout.tv_nsec=10000;
switch (r=sigtimedwait(&s.ss,&info,&timeout)) {
case -1: if (errno!=EAGAIN) error("sigtimedwait");
case SIGIO: puts("SIGIO queue overflow!"); return 1;
}
if (r==signum) handle_io(info.si_fd,info.si_band);
}

info.si_band is identical to pollfd.revents.



Disadvantages of SIGIO

SIGIO has disadvantes, too. The kernel has an event queue, and the events
are delivered in order. The kernel may signal an event on a file descriptor you
already closed.
The queue has a fixed length, and you can’t say how large you need it to be.
If the queue is full, you get a SIGIO signal and no events at all.
You are then expected to flush the queue (by setting the handler for
signum to SIG_DFL) and get the events using poll.
This API reduces the memory traffic, but now you have one syscall per
event. Syscalls are cheap on Linux, but not free.
Handling queue overflows is very annoying. One of the problems with poll
is that you don’t want to maintain the pollfd array.



What’s wrong with the pollfd array?

We write a web server. poll tells us to read from fd 5. read tells us the
connection was dropped. We now close the descriptor and remove the record
from the array.

How do we do that? We can’t just set the events in the record to 0, because
poll will then abort with EBADF and set revents to POLLNVAL. So we need to
copy the last descriptor in the array over this one and reduce the array size.

Now entry 5 in the array is no longer the entry for file descriptor 5! So we
need an extra index array to find the pollfd for a given descriptor. This index
needs to be shrunk and grown and maintained. That is ugly and error prone.

And it is perfectly superfluous.



/dev/poll

Sun added a new poll style API to Solaris a few years ago. You open the
device /dev/poll, and then write the pollfds to that device. You wait for
events with ioctl.
The ioctl then returns how many events are there, and this many pollfds
you can then read from the device. The device only gives you pollfds which
had actual activity.
This means you don’t have to scan the array for the ones with events.
There were several approaches to create such a device for Linux. None of it
made it into the kernel proper.
Oh yes, and the patches were unstable under high load.



/dev/epoll

There is a patch for Linux 2.4 that adds a /dev/epoll device.

int epollfd=open("/dev/misc/eventpoll",O_RDWR);
char* map;
ioctl(epollfd,EP_ALLOC,maxfds); /* hint: number of descriptors */
map=mmap(0, EP_MAP_SIZE(maxfds), PROT_READ, MAP_PRIVATE, epollfd, 0);

You signal interest in an event by writing the pollfd to the device. You
signal disinterest by writing a pollfd with events==0 to the device.



/dev/epoll

The events are fetched with an ioctl:

struct evpoll evp;
for (;;) {
int n;
evp.ep_timeout=1000;
evp.ep_resoff=0;
n=ioctl(e.fd,EP_POLL,&evp);
pfds=(struct pollfd*)(e.map+evp.ep_resoff);
/* now you have n pollfds with events in pfds */
}

Because of mmap, there is no actual copying between kernel and user space.



/dev/epoll

The disadvantage of /dev/epoll is that it is just a patch. Linus does not like
new pseudo devices in the kernel.

He says we already have a dispatcher for the syscalls in the kernel, so if we
want to add something, we add a syscall and not a device or an ioctl.

So the author of /dev/epoll redid the API using syscalls. This API is was
merged into Linux 2.5 (since 2.5.51 even in the documented form ;-) ).

It is now the recommended API for event notiﬁcation in Linux 2.6.



epoll

int epollfd=epoll_create(maxfds);
struct epoll_event x;
x.events=EPOLLIN|EPOLLERR;
x.data.ptr=whatever; /* you can put some cookie here */
epoll_ctl(epollfd,EPOLL_CTL_ADD,fd,&x);
/* changing is analogous: */
epoll_ctl(epollfd,EPOLL_CTL_MOD,fd,&x);
/* deleting -- only x.fd has to be set */
epoll_ctl(epollfd,EPOLL_CTL_DEL,fd,&x);

The EPOLLIN,... constants have the same value as POLLIN,... but the author
wanted to keep all options open. epoll started out edge triggered but is now
level triggered. Add |EPOLLET to switch back.



epoll

This is how you fetch the events:

for (;;) {
struct epoll_event x[100];
int n=epoll_wait(epollfd,x,100,1000); /* 1000 milliseconds */
/* x[0] .. x[n-1] are the events */
}

Note that epoll_event does not contain the actual descriptor!

That’s why the API has a cookie. You can put the ﬁle descriptor there,
but you can also put a pointer to some struct with the context data for this
connection there.



FreeBSD: kqueue

kqueue is a cross between epoll and SIGIO. Like epoll you can have
edge or level triggered events. kqueue can also do ﬁle and directory status
notiﬁcation. The problem is: the API is unwieldly and not well documented.

kqueue is older than epoll. I think Linux should simply have implemented
the kqueue API instead of inventing epoll, but the Linux people insist on doing
all the mistakes of the other people again. For example, the epoll guy initially
thought he could get away without level triggering.

The performance of epoll and kqueue is very similiar.

kqueue is also implemented in OpenBSD, but not in NetBSD.



FreeBSD: kqueue

This is how you signal interest or disinterest in an event:
#include <sys/types.h>
#include <sys/event.h>
#include <sys/time.h>
int kq=kqueue();
struct timespec ts;
EV_SET(&kev, fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
ts.tv_sec=0; ts.tv_nsec=0;
kevent(kq, &kev, 1, 0, 0, &ts); /* I want to read */

EV_SET(&kev, fd, EVFILT_READ, EV_DELETE, 0, 0, 0);
ts.tv_sec=0; ts.tv_nsec=0;
kevent(kq, &kev, 1, 0, 0, &ts); /* I’m done reading */



FreeBSD: kqueue

Here is how you fetch the events:

struct kevent ev[100];
struct timespec ts;
ts.tv_sec=milliseconds/1000;
ts.tv_nsec=(milliseconds%1000)*1000000;
if ((n=kevent(io_master,0,0,y,100,milliseconds!=-1?&ts:0))==-1)
return -1;
for (i=0; i<n; ++i) {
if (ev[i].filter==EVFILT_READ) can_read(ev[i].ident);
if (ev[i].filter==EVFILT_WRITE) can_write(ev[i].ident);
}



Windoze: Completion Ports

Even Microsoft needs to offer something in this ﬁeld.

The Microsoft solution is a thread pool, each thread running something
SIGIO-like mechanism.

This combines the disadvantages of threads with the disadvantages of
SIGIO. It is amazing how Microsoft’s marketing machinery managed to blow
this up into an epic innovation.

I found this quote in the documentation: ”First of all, threads are system
resources that are neither unlimited nor cheap.”

Huh? Not cheap? I thought that was the reason for their existance!



Other reasons for high latency

The POSIX API says that the kernel always has to use the lowest unused
file descriptor when creating a new one (i.e. when accepting a connection,
opening a file, creating a socket).

There is no known way to do this in O(1). Linux uses bit vectors.

The Linux kernel mailing list has carried numerous discussions about
adding an open flag allowing the kernel to return a higher file descriptor, but
it hasn’t happened yet. And it wouldn’t help for sockets.

For Linux and all BSDs this scales well. Here is a graph:



Other reasons for high latency

There is a similar problem for clients and proxies.

When you open a client socket to connect somewhere, and you don’t
explicitly choose a port, the kernel has to choose a free one for you.

No standard says the kernel has to return the lowest port, but this is often
overlooked as problem.

My tests turned out that there is no signiﬁcant difference between IPv4 and
IPv6 latencies on any system, so I only include a graph for IPv4:



Fragmentation and seeking

Hard disks are fast when reading sequentially, but slow when they have to
move the head (i.e. seek).

This is why big server irons use hard disks with 10k or even 15k rpm.

Disk throughput has ceased to be a major concern quite a while ago.

Here is a graph. The hard disk can do 25 MB/sec, i.e. more than twice the
Fast Ethernet capacity. I started a competing download limited to 5 MB/sec on
loopback, for another ﬁle. The throughput for the Ethernet download halves.

The graph also shows how ﬁle fragmentation can slow down read access.



Downloading an ISO image
12000

10000

8000
throughput in K/sec

6000

4000

2000

unfragmented
fragmented
with competing download
0
0 50 100 150 200 250 300 350 400
megabytes transferred



sendfile

sendfile() is like write(), only directly from a file descriptor to a
socket, i.e. no buffer and no read().
That eliminates copying the file data from the buffer cache to a buffer in
user space.
Current NICs (all gigabit ethernet NICs) can do scatter-gather I/O, i.e. they
can take the packet header from a kernel buffer, but the packet contents from
the buffer cache (note: csum partial copy from user).
The result is called Zero Copy TCP and it is the ultimate goal.
Linux and FreeBSD have (different) sendfile syscalls.
NetBSD and OpenBSD don’t have sendfile.



Memory Mapped I/O

Instead of reading from a file and then doing something read-only with the
file contents in the buffer, it is also possible to map the file into user space
memory.

The syscall for this is called mmap, and it is very important for scalable
network I/O, because it eliminates the buffer.

The OS can put that saved memory to good use in the buffer cache.

However, maintaining the page tables is difficult to get to scale properly.
This will be even more important on 64-bit machines. Here are a few graphs:



Filesystem latency

This is usually not a problem these days, but there are pathological cases.

I once ran an FTP server with a big /incoming at a LAN party.

Someone decided to upload his whole porn collection as single JPEG files.

I noticed and kicked him out after 50.000 files, because the server had
become very sluggish.

The OS spent close to 100% CPU time in the kernel, traversing the in-
memory list of files in that directory over and over again.

These days we have advanced file systems like XFS and reiserfs specializing
in this case, and there is directory hashing support for ext3 and FreeBSD UFS.



Asynchronous I/O

POSIX specifies an API for asynchronous I/O. Unfortunately almost nobody
implements it. Linux doesn’t, for example.

glibc has an emulation that creates one thread for each request. This is
much worse than not having the API in the first place, so nobody uses the API.

The idea is to queue a read request. Later, you can then ask the OS if it has
finished the request in the mean time. Or, you can get a signal when the kernel
is done.

The problem is: you have no idea which of the queued requests was finished
when you get the signal.



Asynchronous I/O

Async I/O is meant for ﬁles, not sockets. If you want to read 1000 blocks
from a terabyte database, and you use lseek and read, you force the OS to
read the data in the speciﬁed order. That forces the path of the hard disk head.

With async I/O, the system can sort the blocks so that the hard disk head
will move only once over the whole disk. In principle, this is wonderful. In
practice, it’s worthless.

Solaris, for example, has their own proprietary async I/O API. They also
offer the POSIX API, only to return ENOSYS for all calls.

The most important OS offering async I/O is FreeBSD.



writev, TCP CORK and TCP NOPUSH

An HTTP server writes a header and then the contents of a file to a socket.
If you call write (and then sendfile), the kernel sends one TCP packet for
the header and starts a fresh one for the body.

If the file is only 100 bytes large, it would have fit into one packet.

Obviously this can be solved by having a buffer and copying both the
header and the file contents there. But we want zero-copy TCP, remember?

There are four solutions: writev, TCP_CORK (Linux), TCP_NOPUSH (BSD)
and the FreeBSD sendfile.



writev

writev is like a batch write. You give it a vector of pointers and lengths, and
it writes them all out.

The difference is too small to notice normally, except for TCP connections.

struct iovec x[2];
x[0].iov_base=header; x[0].iov_len=strlen(header);
x[1].iov_base=mapped_file; x[1].iov_len=file_size;
writev(sock,x,2); /* returns bytes written */



TCP CORK

int null=0, eins=1;

/* put cork in bottle */
setsockopt(sock,IPPROTO_TCP,TCP_CORK,&eins,sizeof(eins));
write(sock,header,strlen(header));
write(sock,mapped_file,file_size);
/* pull cork out of bottle */
setsockopt(sock,IPPROTO_TCP,TCP_CORK,&null,sizeof(null));

TCP NOPUSH from BSD does the same, but you have to set the flag to zero
before writing the last write, which is a little awkward.

The sendfile from FreeBSD is like the Linux sendfile plus one writev style
vector for headers and one for footers, so you don’t need TCP NOPUSH.



vfork

fork is very fast on modern Unices, because no actual memory contents is
copied, only the page tables.

However, this can still be comparatively expensive, especially if you were
going to exec a CGI program anyway. That’s why Linux and the BSDs have
implemented a traditional vfork again.

But vfork is not universally faster! On Linux 2.6, vfork is actually slower
(250 ms vs. 180 ms) than fork if the forking program is a statically linked
dietlibc program. For a small glibc program, the difference is 250 ms vs. 320
ms.



So which OS is the best one?

I recommend Linux 2.6. Linux scales O(1) in all benchmarks.

FreeBSD 5.1 is a close second. It is O(1) in all benchmarks except mmap.

Linux 2.4 scales badly for mmap and many processes, use 2.6 instead!

NetBSD does not have kqueue or sendﬁle, only poll. However, it is still a
high performance operating system.

I found OpenBSD thoroughly disappointing. The disk performance sucks,
OpenBSD can’t even serve a single ﬁle with sustained 11 MB/sec on Fast
Ethernet. And they deliberately broke their IPv6 stack.



So which OS is the best one?

From a stability point of view, Linux and NetBSD worked stable all the
time, FreeBSD 5.1-RELEASE panicked under load (that went away with 5.1-
CURRENT) and OpenBSD crashed and panicked even in 3.4-CURRENT. OpenBSD
also surprised me with ”interesting” syslog messages like ”/bsd: full”.

Here are some more ”big picture” graphs. I’ll start with the most surprising
result I encountered during my tests.



HTTP Request latency (without connect time)
6000
FreeBSD 5.1 (kqueue)
Linux 2.4 (sigio)
Linux 2.6 (epoll)
OpenBSD 3.4 (kqueue)
5000

4000
microseconds

3000

2000

1000

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
open connections



HTTP Request latency (without connect time)
25000
FreeBSD 5.1 (kqueue)
Linux 2.4 (sigio)
Linux 2.6 (epoll)
OpenBSD 3.4 (kqueue)
NetBSD 1.6.1 (poll)
20000

15000
microseconds

10000

5000

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
open connections



Questions

Thanks for sitting through this with me in this ungodly hour!

You can get the source code for all the benchmarks via anonymous cvs:

% cvs -d:pserver:cvs@cvs.fefe.de:/cvs -z9 co libowfat
% cvs -d:pserver:cvs@cvs.fefe.de:/cvs -z9 co gatling

My web page is at http://www.fefe.de/.

You can email me at felix-linuxkongress@fefe.de.


2003 scalable networking - unknown

More Related Content

What's hot (20)

Viewers also liked (7)

Similar to 2003 scalable networking - unknown (20)

More from George Ang (20)

2003 scalable networking - unknown