httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Bill Stoddard <b...@wstoddard.com>
Subject Re: Regarding worker MPM and queue_push/pop
Date Thu, 04 Dec 2003 19:40:09 GMT
MATHIHALLI,MADHUSUDAN (HP-Cupertino,ex1) wrote:
> I sure would do that sometime today. The leader probably uses some
> apr_atomic stuff - and I'm trying to see if I can use IA64 native code to do
> the atomics.
> 
> For people at ease with visual stuff, here's the CPU performance that I'm
> getting with worker MPM and SPECweb99_SSL.
> Something looks is terribly wrong.
> 
> httpd.conf:
> ------------
> MPM: worker
> 
> StartServers           1
> MaxClients          2000
> MinSpareThreads      500
> MaxSpareThreads     2000
> ThreadsPerChild      500
> 
> -Madhu
> 
> 

Last I checked (year+ ago), worker performed a bit better on my single CPU AIX machine than
leader-follower. I 
expect l-f could be made to perform better than worker.

It's difficult to tell if your CPU results are fooness or not.  If you have many clients sleeping
on a k-a 
connections, you could easily see CPU usage like this. Are you using CGI scripts? (an aside...
if so better be 
using mod_cgid rather than mod_cgi with worker).  Jeff may have already pointed out to you
a "feature" in the 
AIX that would keep threads hanging around an unusual amount of time waiting for notification
from the kernel 
that the CGI script was dead. Here is the fix for AIX:

http://cvs.apache.org/viewcvs.cgi/httpd-2.0/modules/generators/mod_cgid.c.diff?r1=1.158&r2=1.159

If your specweb99 runs use lots of CGI scripts, you could be hitting this or a variation.

Bill


Mime
View raw message