Running a High-Performance Web Server for BSD

Like other OS's, the listen queue is often the first limit hit. The following are comments from "Aaron Gifford <agifford@InfoWest.COM>" on how to fix this on BSDI 1.x, 2.x, and FreeBSD 2.0 (and earlier):

Edit the following two files:

/usr/include/sys/socket.h /usr/src/sys/sys/socket.h

In each file, look for the following:


    /*
     * Maximum queue length specifiable by listen.
     */
    #define SOMAXCONN       5

Just change the "5" to whatever appears to work. I bumped the two machines I was having problems with up to 32 and haven't noticed the problem since.

After the edit, recompile the kernel and recompile the Apache server then reboot.

FreeBSD 2.1 seems to be perfectly happy, with SOMAXCONN set to 32 already.

Addendum for very heavily loaded BSD servers
from Chuck Murcko <chuck@telebase.com>

If you're running a really busy BSD Apache server, the following are useful things to do if the system is acting sluggish:

Run vmstat to check memory usage, page/swap rates, etc.
Run netstat -m to check mbuf usage
Run fstat to check file descriptor usage

These utilities give you an idea what you'll need to tune in your kernel, and whether it'll help to buy more RAM. Here are some BSD kernel config parameters (actually BSDI, but pertinent to FreeBSD and other 4.4-lite derivatives) from a system getting heavy usage. The tools mentioned above were used, and the system memory was increased to 48 MB before these tuneups. Other system parameters remained unchanged.

maxusers 256 Maxusers drives a lot of other kernel parameters:

Maximum # of processes
Maximum # of processes per user
System wide open files limit
Per-process open files limit
Maximum # of mbuf clusters
Proc/pgrp hash table size

The actual formulae for these derived parameters are in /usr/src/sys/conf/param.c. These calculated parameters can also be overridden (in part) by specifying your own values in the kernel configuration file:


# Network options. NMBCLUSTERS defines the number of mbuf clusters and
# defaults to 256. This machine is a server that handles lots of traffic,
# so we crank that value.
options         SOMAXCONN=256           # max pending connects
options         NMBCLUSTERS=4096        # mbuf clusters at 4096

#
# Misc. options
#
options         CHILD_MAX=512           # maximum number of child processes
options         OPEN_MAX=512            # maximum fds (breaks RPC svcs)

SOMAXCONN is not derived from maxusers, so you'll always need to increase that yourself. We used a value guaranteed to be larger than Apache's default for the listen() of 128, currently.

In many cases, NMBCLUSTERS must be set much larger than would appear necessary at first glance. The reason for this is that if the browser disconnects in mid-transfer, the socket fd associated with that particular connection ends up in the TIME_WAIT state for several minutes, during which time its mbufs are not yet freed. Another reason is that, on server timeouts, some connections end up in FIN_WAIT_2 state forever, because this state doesn't time out on the server, and the browser never sent a final FIN. For more details see the FIN_WAIT_2 page.

Some more info on mbuf clusters (from sys/mbuf.h): /* * Mbufs are of a single size, MSIZE (machine/machparam.h), which * includes overhead. An mbuf may add a single "mbuf cluster" of size * MCLBYTES (also in machine/machparam.h), which has no additional overhead * and is used instead of the internal data area; this is done when * at least MINCLSIZE of data must be stored. */

CHILD_MAX and OPEN_MAX are set to allow up to 512 child processes (different than the maximum value for processes per user ID) and file descriptors. These values may change for your particular configuration (a higher OPEN_MAX value if you've got modules or CGI scripts opening lots of connections or files). If you've got a lot of other activity besides httpd on the same machine, you'll have to set NPROC higher still. In this example, the NPROC value derived from maxusers proved sufficient for our load.

Caveats

Be aware that your system may not boot with a kernel that is configured to use more resources than you have available system RAM. ALWAYS have a known bootable kernel available when tuning your system this way, and use the system tools beforehand to learn if you need to buy more memory before tuning.

RPC services will fail when the value of OPEN_MAX is larger than 256. This is a function of the original implementations of the RPC library, which used a byte value for holding file descriptors. BSDI has partially addressed this limit in its 2.1 release, but a real fix may well await the redesign of RPC itself.

Finally, there's the hard limit of child processes configured in Apache.

For versions of Apache later than 1.0.5 you'll need to change the definition for HARD_SERVER_LIMIT in httpd.h and recompile if you need to run more than the default 150 instances of httpd.

From conf/httpd.conf-dist: # Limit on total number of servers running, i.e., limit on the number # of clients who can simultaneously connect --- if this limit is ever # reached, clients will be LOCKED OUT, so it should NOT BE SET TOO LOW. # It is intended mainly as a brake to keep a runaway server from taking # Unix with it as it spirals down... MaxClients 150 Know what you're doing if you bump this value up, and make sure you've done your system monitoring, RAM expansion, and kernel tuning beforehand. Then you're ready to service some serious hits!

Thanks to Tony Sanders and Chris Torek at BSDI for their helpful suggestions and information.

"M. Teterin" <mi@ALDAN.ziplink.net> writes:

It really does help if your kernel and frequently used utilities are fully optimized. Rebuilding the FreeBSD kernel on an AMD-133 (486-class CPU) web-server with
-m486 -fexpensive-optimizations -fomit-frame-pointer -O2
helped reduce the number of "unable" errors, because the CPU was often maxed out.

More welcome!

If you have tips to contribute, send mail to brian@organic.com