[ejabberd] http_bind & high CPU load: help needed

Stefan Strigler steve at zeank.in-berlin.de
Fri Nov 30 19:17:10 MSK 2007


Hi,

Am Freitag, den 30.11.2007, 17:35 +0300 schrieb George Hazan:
> Hello.
> 
> I need to establish a jabber server with ~10000 connections via http-bind. 
> Unfortunately on my server even 700-800 users create 192% CPU load (two Xeon 
> 3GHz processors). All connections do almost nothing but reading a roster and 
> sending one message in maybe 5 minutes.
> 
> I tried to turn on epoll support (Linux Red Hat Enterprise), but the CPU 
> usage is still too high (140-150%). Postgres database is on another server, 
> and this server does almost nothing, the overall load there is less than 5%.
> 
> Any help is greatly appreciated... I can't set up 10 front-end servers for 
> this task. Ejabberd was compiled from trunk.

Have you tried turning off logging? I haven't done benchmarks yet with
this version of ejabberd. We're using ejabberd-1.1.1 and are able to
handle about 10.000 users per machine (which are less powerful than
yours).

Are you using kernel-poll? (+K true) Dunno if that's the same as epoll.
Are you having support for SMP enabled? 

Cheers, Steve



More information about the ejabberd mailing list