Re: mod_glusterfs (was Re: [Gluster-devel] Hammering start() calls)

[Sascha Ottolski]

Am Sonntag 06 April 2008 20:51:05 schrieb Hannes Dorbath:
> That sounds great, however I'm currently bound to Lighttpd.
>
> I've setup a clean test case to benchmark the issues I'm seeing. With
> -e 600 -a 600 it seems I'm no longer bound to stat(), but to open()
> now.
>
> Following is a trace of a "Hello World" PHP script (FCGI) as well as
> my current configuration.
>
> I think I messed something with iothreads. Without GlusterFS total
> throughput increases a lot with multiple HTTP clients, with GlusterFS
> it decreases. I could not find any load where using iothreads
> improved my situation, so there still seems to be a single point of
> serialisation.
>
> If you have any suggestion on how to improve this configuration, I'd
> really appreciative it.

Hannes,

if you have the time, you might wanna check the list archive for some 
posting of mine, which unfortunately doesn't give you a solution, but 
may convince you that you're not alone :-) In my experiments, I 
couldn't get much more than 900 requests/second pulled from my cluster 
(being small image files, 5-50 KB in size).

but what might be of interest for your setup: I had much worse results 
when testing with lighttpd or nginx (only about 350 req/sec). however, 
with apache-1.3 or apache-2.2, I could oberserve a huge boost when 
enabling the io-cache, but no effect at all for lighttpd and nginx. 
and, as you describe, I never could measure any differences when 
playing with the other translators or start options.

may be you could do your strace for an apache setup also? I guess this 
could shed some light on the reasons that apache works so much better 
(but still way to slow :-() than the modern competetion...


Cheers, Sascha