FreeBSD: stressing 4.2 and Apache with ab (fwd)

[L. Cranswick]

> >   socket: "No buffer space available"
> 
> Sounds (if I remember correctly) that you are running out of NMB
> clusters. Increase them (probably need to add the setting: options
> NMBCLUSTERS="10000") in your kernel config file, or possibly by sysctl
> (kern.ipc.nmbclusters)
> 
> -Steven

Thanks for this, I have done an incoherant  draft page on making 
this happen at (corrections, insults, etc appreciated):
Optimising FreeBSD 4.2 UNIX and Apache for High Performance 
PC Webserving
 http://www.ccp14.ac.uk/solution/bsdunix/optimise_freebsd_for_apache.html

Things are now much more robust with respect to stress testing
using ab (apachebench).

A few more silly questions (telling me to not worry about
it and get a life is an acceptable answer).

On web searching:
I cannot find anything about tuning FreeBSD with Apache though
there are Linux equivalents - what are the parameters that
should be tweaked and what are the implications - especially
if RAM might be becoming tight - will it just swap to disk?

-----

With nmbclusters = 10000 - it now takes a lot more to drag
FreeBSD down in these tests.  (is still a 300 MHz Pentium
with 128 Meg RAM)

But when I do:
  
  /usr/local/apache/bin/ab -n 10000 -c 1000 http://system.dl.ac.uk/test/test.pl

I get:
  Benchmarking system.dl.ac.uk (be patient)...
    socket: Too many open files in system 

And the FreeBSD consol/logs reports: /kernel: file: table is full
/var/run/utmp: Too many open files in system.
no memory for rx list, etc.

Does this give suggestions on parameters to tweak 
(just increase nmbclusters more) or am
I asking too much from the system?  I would like to
get a strategy that FreeBSD/apache starts to go slow 
rather than barfs when it starts to get under
excessive load - if this is a reasonable request.  
Or at least get the max out of it before FreeBSD starts
barfing.

(I should note that FreeBSD even in default mode is
majorly out-performing my SGI O2 on these benchmarks
and seems to be doing better than the computer mentioned
in 
 http://perl.apache.org/guide/performance.html#Tuning_with_the_crashme_Script

Cheers,

Lachlan.


-- 
Lachlan M. D. Cranswick

Collaborative Computational Project No 14 (CCP14)
    for Single Crystal and Powder Diffraction
Daresbury Laboratory, Warrington, WA4 4AD U.K
Tel: +44-1925-603703  Fax: +44-1925-603124
E-mail: l.cranswick at dl.ac.uk  Ext: 3703  Room C14
                           http://www.ccp14.ac.uk