Apache Performance Notes

Author: Dean Gaudet

Introduction

Apache is a general webserver, which is designed to be correct first, and fast second. Even so, its performance is quite satisfactory. Most sites have less than 10Mbits of outgoing bandwidth, which Apache can fill using only a low end Pentium-based webserver. In practice sites with more bandwidth require more than one machine to fill the bandwidth due to other constraints (such as CGI or database transaction overhead). For these reasons the development focus has been mostly on correctness and configurability.

Unfortunately many folks overlook these facts and cite raw performance numbers as if they are some indication of the quality of a web server product. There is a bare minimum performance that is acceptable, beyond that extra speed only caters to a much smaller segment of the market. But in order to avoid this hurdle to the acceptance of Apache in some markets, effort was put into Apache 1.3 to bring performance up to a point where the difference with other high-end webservers is minimal.

Finally there are the folks who just plain want to see how fast something can go. The author falls into this category. The rest of this document is dedicated to these folks who want to squeeze every last bit of performance out of Apache's current model, and want to understand why it does some things which slow it down.

Note that this is tailored towards Apache 1.3 on Unix. Some of it applies to Apache on NT. Apache on NT has not been tuned for performance yet, in fact it probably performs very poorly because NT performance requires a different programming model.

Hardware and Operating System Issues

The single biggest hardware issue affecting webserver performance is RAM. A webserver should never ever have to swap, swapping increases the latency of each request beyond a point that users consider "fast enough". This causes users to hit stop and reload, further increasing the load. You can, and should, control the MaxClients setting so that your server does not spawn so many children it starts swapping.

Beyond that the rest is mundane: get a fast enough CPU, a fast enough network card, and fast enough disks, where "fast enough" is something that needs to be determined by experimentation.

Operating system choice is largely a matter of local concerns. But a general guideline is to always apply the latest vendor TCP/IP patches. HTTP serving completely breaks many of the assumptions built into Unix kernels up through 1994 and even 1995. Good choices include recent FreeBSD, and Linux.

Run-Time Configuration Issues

HostnameLookups

Prior to Apache 1.3, HostnameLookups defaulted to On. This adds latency to every request because it requires a DNS lookup to complete before the request is finished. In Apache 1.3 this setting defaults to Off. However (1.3 or later), if you use any Allow from domain or Deny from domain directives then you will pay for a double reverse DNS lookup (a reverse, followed by a forward to make sure that the reverse is not being spoofed). So for the highest performance avoid using these directives (it's fine to use IP addresses rather than domain names).

Note that it's possible to scope the directives, such as within a <Location /server-status> section. In this case the DNS lookups are only performed on requests matching the criteria. Here's an example which disables lookups except for .html and .cgi files:

HostnameLookups off
<Files ~ "\.(html|cgi)$>
    HostnameLookups on
</Files>

But even still, if you just need DNS names in some CGIs you could consider doing the gethostbyname call in the specific CGIs that need it.

FollowSymLinks and SymLinksIfOwnerMatch

Wherever in your URL-space you do not have an Options FollowSymLinks, or you do have an Options SymLinksIfOwnerMatch Apache will have to issue extra system calls to check up on symlinks. One extra call per filename component. For example, if you had:

DocumentRoot /www/htdocs
<Directory />
    Options SymLinksIfOwnerMatch
</Directory>

and a request is made for the URI /index.html. Then Apache will perform lstat(2) on /www, /www/htdocs, and /www/htdocs/index.html. The results of these lstats are never cached, so they will occur on every single request. If you really desire the symlinks security checking you can do something like this:

DocumentRoot /www/htdocs
<Directory />
    Options FollowSymLinks
</Directory>
<Directory /www/htdocs>
    Options -FollowSymLinks +SymLinksIfOwnerMatch
</Directory>

This at least avoids the extra checks for the DocumentRoot path. Note that you'll need to add similar sections if you have any Alias or RewriteRule paths outside of your document root. For highest performance, and no symlink protection, set FollowSymLinks everywhere, and never set SymLinksIfOwnerMatch.

AllowOverride

Wherever in your URL-space you allow overrides (typically .htaccess files) Apache will attempt to open .htaccess for each filename component. For example,

DocumentRoot /www/htdocs
<Directory />
    AllowOverride all
</Directory>

and a request is made for the URI /index.html. Then Apache will attempt to open /.htaccess, /www/.htaccess, and /www/htdocs/.htaccess. The solutions are similar to the previous case of

Options
FollowSymLinks

. For highest performance use AllowOverride None everywhere in your filesystem.

Negotiation

If at all possible, avoid content-negotiation if you're really interested in every last ounce of performance. In practice the benefits of negotiation outweigh the performance penalties. There's one case where you can speed up the server. Instead of using a wildcard such as:

DirectoryIndex index

Use a complete list of options:

DirectoryIndex index.cgi index.pl index.shtml index.html

where you list the most common choice first.

Process Creation

Prior to Apache 1.3 the MinSpareServers, MaxSpareServers, and StartServers settings all had drastic effects on benchmark results. In particular, Apache required a "ramp-up" period in order to reach a number of children sufficient to serve the load being applied. After the initial spawning of StartServers children, only one child per second would be created to satisfy the MinSpareServers setting. So a server being accessed by 100 simultaneous clients, using the default StartServers of 5 would take on the order 95 seconds to spawn enough children to handle the load. This works fine in practice on real-life servers, because they aren't restarted frequently. But does really poorly on benchmarks which might only run for ten minutes.

The one-per-second rule was implemented in an effort to avoid swamping the machine with the startup of new children. If the machine is busy spawning children it can't service requests. But it has such a drastic effect on the perceived performance of Apache that it had to be replaced. As of Apache 1.3, the code will relax the one-per-second rule. It will spawn one, wait a second, then spawn two, wait a second, then spawn four, and it will continue exponentially until it is spawning 32 children per second. It will stop whenever it satisfies the MinSpareServers setting.

This appears to be responsive enough that it's almost unnecessary to twiddle the MinSpareServers, MaxSpareServers and StartServers knobs. When more than 4 children are spawned per second, a message will be emitted to the ErrorLog. If you see a lot of these errors then consider tuning these settings. Use the mod_status output as a guide.

Related to process creation is process death induced by the MaxRequestsPerChild setting. By default this is 0, which means that there is no limit to the number of requests handled per child. If your configuration currently has this set to some very low number, such as 30, you may want to bump this up significantly. If you are running SunOS or an old version of Solaris, limit this to 10000 or so because of memory leaks.

When keep-alives are in use, children will be kept busy doing nothing waiting for more requests on the already open connection. The default KeepAliveTimeout of 15 seconds attempts to minimize this effect. The tradeoff here is between network bandwidth and server resources. In no event should you raise this above about 60 seconds, as most of the benefits are lost.

Compile-Time Configuration Issues

mod_status and ExtendedStatus On

If you include mod_status and you also set ExtendedStatus On when building and running Apache, then on every request Apache will perform two calls to gettimeofday(2) (or times(2) depending on your operating system), and (pre-1.3) several extra calls to time(2). This is all done so that the status report contains timing indications. For highest performance, set ExtendedStatus off (which is the default).

accept Serialization - multiple sockets

This discusses a shortcoming in the Unix socket API. Suppose your web server uses multiple Listen statements to listen on either multiple ports or multiple addresses. In order to test each socket to see if a connection is ready Apache uses select(2). select(2) indicates that a socket has zero or at least one connection waiting on it. Apache's model includes multiple children, and all the idle ones test for new connections at the same time. A naive implementation looks something like this (these examples do not match the code, they're contrived for pedagogical purposes):

    for (;;) {
	for (;;) {
	    fd_set accept_fds;

	    FD_ZERO (&accept_fds);
	    for (i = first_socket; i <= last_socket; ++i) {
		FD_SET (i, &accept_fds);
	    }
	    rc = select (last_socket+1, &accept_fds, NULL, NULL, NULL);
	    if (rc < 1) continue;
	    new_connection = -1;
	    for (i = first_socket; i <= last_socket; ++i) {
		if (FD_ISSET (i, &accept_fds)) {
		    new_connection = accept (i, NULL, NULL);
		    if (new_connection != -1) break;
		}
	    }
	    if (new_connection != -1) break;
	}
	process the new_connection;
    }

But this naive implementation has a serious starvation problem. Recall that multiple children execute this loop at the same time, and so multiple children will block at select when they are in between requests. All those blocked children will awaken and return from select when a single request appears on any socket (the number of children which awaken varies depending on the operating system and timing issues). They will all then fall down into the loop and try to accept the connection. But only one will succeed (assuming there's still only one connection ready), the rest will be blocked in accept. This effectively locks those children into serving requests from that one socket and no other sockets, and they'll be stuck there until enough new requests appear on that socket to wake them all up. This starvation problem was first documented in PR#467. There are at least two solutions.

One solution is to make the sockets non-blocking. In this case the accept won't block the children, and they will be allowed to continue immediately. But this wastes CPU time. Suppose you have ten idle children in select, and one connection arrives. Then nine of those children will wake up, try to accept the connection, fail, and loop back into select, accomplishing nothing. Meanwhile none of those children are servicing requests that occurred on other sockets until they get back up to the select again. Overall this solution does not seem very fruitful unless you have as many idle CPUs (in a multiprocessor box) as you have idle children, not a very likely situation.

Another solution, the one used by Apache, is to serialize entry into the inner loop. The loop looks like this (differences highlighted):

    for (;;) {
	accept_mutex_on ();
	for (;;) {
	    fd_set accept_fds;

	    FD_ZERO (&accept_fds);
	    for (i = first_socket; i <= last_socket; ++i) {
		FD_SET (i, &accept_fds);
	    }
	    rc = select (last_socket+1, &accept_fds, NULL, NULL, NULL);
	    if (rc < 1) continue;
	    new_connection = -1;
	    for (i = first_socket; i <= last_socket; ++i) {
		if (FD_ISSET (i, &accept_fds)) {
		    new_connection = accept (i, NULL, NULL);
		    if (new_connection != -1) break;
		}
	    }
	    if (new_connection != -1) break;
	}
	accept_mutex_off ();
	process the new_connection;
    }

The functions accept_mutex_on and accept_mutex_off implement a mutual exclusion semaphore. Only one child can have the mutex at any time. There are several choices for implementing these mutexes. The choice is defined in src/conf.h (pre-1.3) or src/include/ap_config.h (1.3 or later). Some architectures do not have any locking choice made, on these architectures it is unsafe to use multiple Listen directives.

USE_FLOCK_SERIALIZED_ACCEPT: This method uses the flock(2) system call to lock a lock file (located by the LockFile directive).
USE_FCNTL_SERIALIZED_ACCEPT: This method uses the fcntl(2) system call to lock a lock file (located by the LockFile directive).
USE_SYSVSEM_SERIALIZED_ACCEPT: (1.3 or later) This method uses SysV-style semaphores to implement the mutex. Unfortunately SysV-style semaphores have some bad side-effects. One is that it's possible Apache will die without cleaning up the semaphore (see the ipcs(8) man page). The other is that the semaphore API allows for a denial of service attack by any CGIs running under the same uid as the webserver (i.e., all CGIs, unless you use something like suexec or cgiwrapper). For these reasons this method is not used on any architecture except IRIX (where the previous two are prohibitively expensive on most IRIX boxes).
USE_USLOCK_SERIALIZED_ACCEPT: (1.3 or later) This method is only available on IRIX, and uses usconfig(2) to create a mutex. While this method avoids the hassles of SysV-style semaphores, it is not the default for IRIX. This is because on single processor IRIX boxes (5.3 or 6.2) the uslock code is two orders of magnitude slower than the SysV-semaphore code. On multi-processor IRIX boxes the uslock code is an order of magnitude faster than the SysV-semaphore code. Kind of a messed up situation. So if you're using a multiprocessor IRIX box then you should rebuild your webserver with -DUSE_USLOCK_SERIALIZED_ACCEPT on the EXTRA_CFLAGS.
USE_PTHREAD_SERIALIZED_ACCEPT: (1.3 or later) This method uses POSIX mutexes and should work on any architecture implementing the full POSIX threads specification, however appears to only work on Solaris (2.5 or later), and even then only in certain configurations. If you experiment with this you should watch out for your server hanging and not responding. Static content only servers may work just fine.

If your system has another method of serialization which isn't in the above list then it may be worthwhile adding code for it (and submitting a patch back to Apache).

Another solution that has been considered but never implemented is to partially serialize the loop -- that is, let in a certain number of processes. This would only be of interest on multiprocessor boxes where it's possible multiple children could run simultaneously, and the serialization actually doesn't take advantage of the full bandwidth. This is a possible area of future investigation, but priority remains low because highly parallel web servers are not the norm.

Ideally you should run servers without multiple Listen statements if you want the highest performance. But read on.

accept Serialization - single socket

The above is fine and dandy for multiple socket servers, but what about single socket servers? In theory they shouldn't experience any of these same problems because all children can just block in accept(2) until a connection arrives, and no starvation results. In practice this hides almost the same "spinning" behaviour discussed above in the non-blocking solution. The way that most TCP stacks are implemented, the kernel actually wakes up all processes blocked in accept when a single connection arrives. One of those processes gets the connection and returns to user-space, the rest spin in the kernel and go back to sleep when they discover there's no connection for them. This spinning is hidden from the user-land code, but it's there nonetheless. This can result in the same load-spiking wasteful behaviour that a non-blocking solution to the multiple sockets case can.

For this reason we have found that many architectures behave more "nicely" if we serialize even the single socket case. So this is actually the default in almost all cases. Crude experiments under Linux (2.0.30 on a dual Pentium pro 166 w/128Mb RAM) have shown that the serialization of the single socket case causes less than a 3% decrease in requests per second over unserialized single-socket. But unserialized single-socket showed an extra 100ms latency on each request. This latency is probably a wash on long haul lines, and only an issue on LANs. If you want to override the single socket serialization you can define SINGLE_LISTEN_UNSERIALIZED_ACCEPT and then single-socket servers will not serialize at all.

Lingering Close

As discussed in draft-ietf-http-connection-00.txt section 8, in order for an HTTP server to reliably implement the protocol it needs to shutdown each direction of the communication independently (recall that a TCP connection is bi-directional, each half is independent of the other). This fact is often overlooked by other servers, but is correctly implemented in Apache as of 1.2.

When this feature was added to Apache it caused a flurry of problems on various versions of Unix because of a shortsightedness. The TCP specification does not state that the FIN_WAIT_2 state has a timeout, but it doesn't prohibit it. On systems without the timeout, Apache 1.2 induces many sockets stuck forever in the FIN_WAIT_2 state. In many cases this can be avoided by simply upgrading to the latest TCP/IP patches supplied by the vendor. In cases where the vendor has never released patches (i.e., SunOS4 -- although folks with a source license can patch it themselves) we have decided to disable this feature.

There are two ways of accomplishing this. One is the socket option SO_LINGER. But as fate would have it, this has never been implemented properly in most TCP/IP stacks. Even on those stacks with a proper implementation (i.e., Linux 2.0.31) this method proves to be more expensive (cputime) than the next solution.

For the most part, Apache implements this in a function called lingering_close (in http_main.c). The function looks roughly like this:

    void lingering_close (int s)
    {
	char junk_buffer[2048];

	/* shutdown the sending side */
	shutdown (s, 1);

	signal (SIGALRM, lingering_death);
	alarm (30);

	for (;;) {
	    select (s for reading, 2 second timeout);
	    if (error) break;
	    if (s is ready for reading) {
		if (read (s, junk_buffer, sizeof (junk_buffer)) <= 0) {
		    break;
		}
		/* just toss away whatever is read */
	    }
	}

	close (s);
    }

This naturally adds some expense at the end of a connection, but it is required for a reliable implementation. As HTTP/1.1 becomes more prevalent, and all connections are persistent, this expense will be amortized over more requests. If you want to play with fire and disable this feature you can define NO_LINGCLOSE, but this is not recommended at all. In particular, as HTTP/1.1 pipelined persistent connections come into use lingering_close is an absolute necessity (and pipelined connections are faster, so you want to support them).

Scoreboard File

Apache's parent and children communicate with each other through something called the scoreboard. Ideally this should be implemented in shared memory. For those operating systems that we either have access to, or have been given detailed ports for, it typically is implemented using shared memory. The rest default to using an on-disk file. The on-disk file is not only slow, but it is unreliable (and less featured). Peruse the src/main/conf.h file for your architecture and look for either USE_MMAP_SCOREBOARD or USE_SHMGET_SCOREBOARD. Defining one of those two (as well as their companions HAVE_MMAP and HAVE_SHMGET respectively) enables the supplied shared memory code. If your system has another type of shared memory, edit the file src/main/http_main.c and add the hooks necessary to use it in Apache. (Send us back a patch too please.)

Historical note: The Linux port of Apache didn't start to use shared memory until version 1.2 of Apache. This oversight resulted in really poor and unreliable behaviour of earlier versions of Apache on Linux.

`DYNAMIC_MODULE_LIMIT`

If you have no intention of using dynamically loaded modules (you probably don't if you're reading this and tuning your server for every last ounce of performance) then you should add -DDYNAMIC_MODULE_LIMIT=0 when building your server. This will save RAM that's allocated only for supporting dynamically loaded modules.

Appendix: Detailed Analysis of a Trace

Here is a system call trace of Apache 1.3 running on Linux. The run-time configuration file is essentially the default plus:

<Directory />
    AllowOverride none
    Options FollowSymLinks
</Directory>

The file being requested is a static 6K file of no particular content. Traces of non-static requests or requests with content negotiation look wildly different (and quite ugly in some cases). First the entire trace, then we'll examine details. (This was generated by the strace program, other similar programs include truss, ktrace, and par.)

accept(15, {sin_family=AF_INET, sin_port=htons(22283), sin_addr=inet_addr("127.0.0.1")}, [16]) = 3
flock(18, LOCK_UN)                      = 0
sigaction(SIGUSR1, {SIG_IGN}, {0x8059954, [], SA_INTERRUPT}) = 0
getsockname(3, {sin_family=AF_INET, sin_port=htons(8080), sin_addr=inet_addr("127.0.0.1")}, [16]) = 0
setsockopt(3, IPPROTO_TCP1, [1], 4)     = 0
read(3, "GET /6k HTTP/1.0\r\nUser-Agent: "..., 4096) = 60
sigaction(SIGUSR1, {SIG_IGN}, {SIG_IGN}) = 0
time(NULL)                              = 873959960
gettimeofday({873959960, 404935}, NULL) = 0
stat("/home/dgaudet/ap/apachen/htdocs/6k", {st_mode=S_IFREG|0644, st_size=6144, ...}) = 0
open("/home/dgaudet/ap/apachen/htdocs/6k", O_RDONLY) = 4
mmap(0, 6144, PROT_READ, MAP_PRIVATE, 4, 0) = 0x400ee000
writev(3, [{"HTTP/1.1 200 OK\r\nDate: Thu, 11"..., 245}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 6144}], 2) = 6389
close(4)                                = 0
time(NULL)                              = 873959960
write(17, "127.0.0.1 - - [10/Sep/1997:23:39"..., 71) = 71
gettimeofday({873959960, 417742}, NULL) = 0
times({tms_utime=5, tms_stime=0, tms_cutime=0, tms_cstime=0}) = 446747
shutdown(3, 1 /* send */)               = 0
oldselect(4, [3], NULL, [3], {2, 0})    = 1 (in [3], left {2, 0})
read(3, "", 2048)                       = 0
close(3)                                = 0
sigaction(SIGUSR1, {0x8059954, [], SA_INTERRUPT}, {SIG_IGN}) = 0
munmap(0x400ee000, 6144)                = 0
flock(18, LOCK_EX)                      = 0

Notice the accept serialization:

flock(18, LOCK_UN)                      = 0
...
flock(18, LOCK_EX)                      = 0

These two calls can be removed by defining SINGLE_LISTEN_UNSERIALIZED_ACCEPT as described earlier.

Notice the SIGUSR1 manipulation:

sigaction(SIGUSR1, {SIG_IGN}, {0x8059954, [], SA_INTERRUPT}) = 0
...
sigaction(SIGUSR1, {SIG_IGN}, {SIG_IGN}) = 0
...
sigaction(SIGUSR1, {0x8059954, [], SA_INTERRUPT}, {SIG_IGN}) = 0

This is caused by the implementation of graceful restarts. When the parent receives a SIGUSR1 it sends a SIGUSR1 to all of its children (and it also increments a "generation counter" in shared memory). Any children that are idle (between connections) will immediately die off when they receive the signal. Any children that are in keep-alive connections, but are in between requests will die off immediately. But any children that have a connection and are still waiting for the first request will not die off immediately.

To see why this is necessary, consider how a browser reacts to a closed connection. If the connection was a keep-alive connection and the request being serviced was not the first request then the browser will quietly reissue the request on a new connection. It has to do this because the server is always free to close a keep-alive connection in between requests (i.e., due to a timeout or because of a maximum number of requests). But, if the connection is closed before the first response has been received the typical browser will display a "document contains no data" dialogue (or a broken image icon). This is done on the assumption that the server is broken in some way (or maybe too overloaded to respond at all). So Apache tries to avoid ever deliberately closing the connection before it has sent a single response. This is the cause of those SIGUSR1 manipulations.

Note that it is theoretically possible to eliminate all three of these calls. But in rough tests the gain proved to be almost unnoticeable.

In order to implement virtual hosts, Apache needs to know the local socket address used to accept the connection:

getsockname(3, {sin_family=AF_INET, sin_port=htons(8080), sin_addr=inet_addr("127.0.0.1")}, [16]) = 0

It is possible to eliminate this call in many situations (such as when there are no virtual hosts, or when Listen directives are used which do not have wildcard addresses). But no effort has yet been made to do these optimizations.

Apache turns off the Nagle algorithm:

setsockopt(3, IPPROTO_TCP1, [1], 4)     = 0

because of problems described in a paper by John Heidemann.

Notice the two time calls:

time(NULL)                              = 873959960
...
time(NULL)                              = 873959960

One of these occurs at the beginning of the request, and the other occurs as a result of writing the log. At least one of these is required to properly implement the HTTP protocol. The second occurs because the Common Log Format dictates that the log record include a timestamp of the end of the request. A custom logging module could eliminate one of the calls. Or you can use a method which moves the time into shared memory, see the patches section below.

As described earlier, ExtendedStatus On causes two gettimeofday calls and a call to times:

gettimeofday({873959960, 404935}, NULL) = 0
...
gettimeofday({873959960, 417742}, NULL) = 0
times({tms_utime=5, tms_stime=0, tms_cutime=0, tms_cstime=0}) = 446747

These can be removed by setting ExtendedStatus Off (which is the default).

It might seem odd to call stat:

stat("/home/dgaudet/ap/apachen/htdocs/6k", {st_mode=S_IFREG|0644, st_size=6144, ...}) = 0

This is part of the algorithm which calculates the PATH_INFO for use by CGIs. In fact if the request had been for the URI /cgi-bin/printenv/foobar then there would be two calls to stat. The first for /home/dgaudet/ap/apachen/cgi-bin/printenv/foobar which does not exist, and the second for /home/dgaudet/ap/apachen/cgi-bin/printenv, which does exist. Regardless, at least one stat call is necessary when serving static files because the file size and modification times are used to generate HTTP headers (such as Content-Length, Last-Modified) and implement protocol features (such as If-Modified-Since). A somewhat more clever server could avoid the stat when serving non-static files, however doing so in Apache is very difficult given the modular structure.

All static files are served using mmap:

mmap(0, 6144, PROT_READ, MAP_PRIVATE, 4, 0) = 0x400ee000
...
munmap(0x400ee000, 6144)                = 0

On some architectures it's slower to mmap small files than it is to simply read them. The define MMAP_THRESHOLD can be set to the minimum size required before using mmap. By default it's set to 0 (except on SunOS4 where experimentation has shown 8192 to be a better value). Using a tool such as lmbench you can determine the optimal setting for your environment.

You may also wish to experiment with MMAP_SEGMENT_SIZE (default 32768) which determines the maximum number of bytes that will be written at a time from mmap()d files. Apache only resets the client's Timeout in between write()s. So setting this large may lock out low bandwidth clients unless you also increase the Timeout.

It may even be the case that mmap isn't used on your architecture, if so then defining USE_MMAP_FILES and HAVE_MMAP might work (if it works then report back to us).

Apache does its best to avoid copying bytes around in memory. The first write of any request typically is turned into a writev which combines both the headers and the first hunk of data:

writev(3, [{"HTTP/1.1 200 OK\r\nDate: Thu, 11"..., 245}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 6144}], 2) = 6389

When doing HTTP/1.1 chunked encoding Apache will generate up to four element writevs. The goal is to push the byte copying into the kernel, where it typically has to happen anyhow (to assemble network packets). On testing, various Unixes (BSDI 2.x, Solaris 2.5, Linux 2.0.31+) properly combine the elements into network packets. Pre-2.0.31 Linux will not combine, and will create a packet for each element, so upgrading is a good idea. Defining NO_WRITEV will disable this combining, but result in very poor chunked encoding performance.

The log write:

write(17, "127.0.0.1 - - [10/Sep/1997:23:39"..., 71) = 71

can be deferred by defining BUFFERED_LOGS. In this case up to PIPE_BUF bytes (a POSIX defined constant) of log entries are buffered before writing. At no time does it split a log entry across a PIPE_BUF boundary because those writes may not be atomic. (i.e., entries from multiple children could become mixed together). The code does it best to flush this buffer when a child dies.

The lingering close code causes four system calls:

shutdown(3, 1 /* send */)               = 0
oldselect(4, [3], NULL, [3], {2, 0})    = 1 (in [3], left {2, 0})
read(3, "", 2048)                       = 0
close(3)                                = 0

which were described earlier.

Let's apply some of these optimizations: -DSINGLE_LISTEN_UNSERIALIZED_ACCEPT -DBUFFERED_LOGS and ExtendedStatus Off. Here's the final trace:

accept(15, {sin_family=AF_INET, sin_port=htons(22286), sin_addr=inet_addr("127.0.0.1")}, [16]) = 3
sigaction(SIGUSR1, {SIG_IGN}, {0x8058c98, [], SA_INTERRUPT}) = 0
getsockname(3, {sin_family=AF_INET, sin_port=htons(8080), sin_addr=inet_addr("127.0.0.1")}, [16]) = 0
setsockopt(3, IPPROTO_TCP1, [1], 4)     = 0
read(3, "GET /6k HTTP/1.0\r\nUser-Agent: "..., 4096) = 60
sigaction(SIGUSR1, {SIG_IGN}, {SIG_IGN}) = 0
time(NULL)                              = 873961916
stat("/home/dgaudet/ap/apachen/htdocs/6k", {st_mode=S_IFREG|0644, st_size=6144, ...}) = 0
open("/home/dgaudet/ap/apachen/htdocs/6k", O_RDONLY) = 4
mmap(0, 6144, PROT_READ, MAP_PRIVATE, 4, 0) = 0x400e3000
writev(3, [{"HTTP/1.1 200 OK\r\nDate: Thu, 11"..., 245}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 6144}], 2) = 6389
close(4)                                = 0
time(NULL)                              = 873961916
shutdown(3, 1 /* send */)               = 0
oldselect(4, [3], NULL, [3], {2, 0})    = 1 (in [3], left {2, 0})
read(3, "", 2048)                       = 0
close(3)                                = 0
sigaction(SIGUSR1, {0x8058c98, [], SA_INTERRUPT}, {SIG_IGN}) = 0
munmap(0x400e3000, 6144)                = 0

That's 19 system calls, of which 4 remain relatively easy to remove, but don't seem worth the effort.

Appendix: Patches Available

There are several performance patches available for 1.3. Although they may not apply cleanly to the current version, it shouldn't be difficult for someone with a little C knowledge to update them. In particular:

A patch to remove all time(2) system calls.
A patch to remove various system calls from mod_include, these calls are used by few sites but required for backwards compatibility.
A patch which integrates the above two plus a few other speedups at the cost of removing some functionality.

Appendix: The Pre-Forking Model

Apache (on Unix) is a pre-forking model server. The parent process is responsible only for forking child processes, it does not serve any requests or service any network sockets. The child processes actually process connections, they serve multiple connections (one at a time) before dying. The parent spawns new or kills off old children in response to changes in the load on the server (it does so by monitoring a scoreboard which the children keep up to date).

This model for servers offers a robustness that other models do not. In particular, the parent code is very simple, and with a high degree of confidence the parent will continue to do its job without error. The children are complex, and when you add in third party code via modules, you risk segmentation faults and other forms of corruption. Even should such a thing happen, it only affects one connection and the server continues serving requests. The parent quickly replaces the dead child.

Pre-forking is also very portable across dialects of Unix. Historically this has been an important goal for Apache, and it continues to remain so.

The pre-forking model comes under criticism for various performance aspects. Of particular concern are the overhead of forking a process, the overhead of context switches between processes, and the memory overhead of having multiple processes. Furthermore it does not offer as many opportunities for data-caching between requests (such as a pool of mmapped files). Various other models exist and extensive analysis can be found in the papers of the JAWS project. In practice all of these costs vary drastically depending on the operating system.

Apache's core code is already multithread aware, and Apache version 1.3 is multithreaded on NT. There have been at least two other experimental implementations of threaded Apache, one using the 1.3 code base on DCE, and one using a custom user-level threads package and the 1.0 code base; neither is available publically. There is also an experimental port of Apache 1.3 to Netscape's Portable Run Time, which is available (but you're encouraged to join the new-httpd mailing list if you intend to use it). Part of our redesign for version 2.0 of Apache will include abstractions of the server model so that we can continue to support the pre-forking model, and also support various threaded models.