[ejabberd] Performance of ejabberd's BOSH
zcongwei at gmail.com
Thu Sep 10 07:30:25 MSD 2009
2009/9/9 Carlos Abalde <carlos.abalde at gmail.com>
> > http://dev.esl.eu/blog/2009/05/05/benchmarking-bosh-services/
> > (read the comments as well)
> After extensive testing (I'm one of the commentators in the previous
> post) I using ejabberd with the built-in BOSH connection manager. It's
> true that it consumes a good amount of memory, but definitively is the
> best and most scalable option. I have made several modification to he
> base implementation to save some memory, but are too specific to our
> environment and are not generally reusable.
Can you give us any suggestion about the modification?
We're planning to use ejabberd, but many client will use http connection, I
think maybe more than 50% clients will through http to login. and the
clients will be about 500Ks
We are considering the structure of the servers, and how to implement...
> As Badlop suggested I would recommend you using the new http-bind
> implementation in ejabberd. I've been checking it a few days ago and it
> does a much more wise handling of in-coming/out-coming XML messages. I
> also recommend you compare memory consumption with and withoy BOSH
> (5222). Memory consumptions depends a lot of your test (rostes sizes,
> message rates, etc.). Of coure, don't forget enabling smp, kernel poll
> and son on.
In our test now, we only login and send the presence.
All the users have 80 contacts
We also test in 5222 port through socket connection (use Tsung test tool)
It's about more than 200KB memory consumption per connection without BOSH(in
5222 port), this also very wise,
and I saw sombody's test result(http://www.ejabberd.im/node/2840), they got
It also very strange....
We had enable the smp and kernel poll..
But other configuration is by default.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the ejabberd