[ejabberd] Identifying pubsub performance bottlenecks

Matthias Rieber ml-erlang at zu-con.org
Wed Oct 17 00:38:38 MSK 2012


Hi,

> On 15.10.2012 17:17, Matthias Rieber wrote:
> > Are there any news about the ejabberd3 release?
> > http://blog.process-one.net/road-ejabberd-3/ mention October 2012.
>
> AFAIR, ejabberd3 will not give you better pubsub performance...

maybe not in general, but as EJAB-1461 says, pubsub with many 
subscriptions has been improved. When I run the test against ejabberd 
2.1.11 and 3-alpha-5 the difference is tremendous.

The test looks like this:

 connect 10000 User with a rate of 20 per second
 sleep between 1 and 600 seconds
 subscribe 50 random nodes out of 2000
 sleep

There are about 10-20 events/s for these 2000 nodes.

Results
=======

Time for subscription
---------------------

ejabberd2: 5mn 17sec(highest 10sec mean), 38.58 msec(lowest 10sec 
mean), 1mn 10sec(Mean)
ejabberd3: 0.76 sec(highest 10sec mean), 40.21 msec(lowest 10sec mean), 
0.19 sec(Mean)

Here are two graphs of this test:

 http://fex.zu-con.org/fop/UYfy8o2w/ejabberd2.png
 http://fex.zu-con.org/fop/4POip5IH/ejabberd3.png


Ejabberd3 seems to have some issues, like http-bind doesn't work with 
strophe.js or subscriptions can't be listed, but as far as I can see, 
the events are delivered properly.

I guess there are no plans to backport this feature/fix to ejabberd2? 
and well, I already asked for ejabberd3 ;)

Matthias





More information about the ejabberd mailing list