[Server-devel] testing ejabberd

[Douglas Bagnall]

I've written up my recent testing of ejabberd for the wiki:

http://wiki.laptop.org/go/Ejabberd_resource_tests

It is not completely satisfactory: I don't have the resources to test
up to 3000 active users which I believe is an important target.  At
lower numbers, however, ejabberd's memory consumption seems to be
linear, and it looks to be roughly the case that 0.5 GB per 1000 users
is enough.  (Just barely -- that's a limit, not a recommendation).

With 1200 users making some communication every 15 seconds, the 2GHz
dual core pentium was bouncing along with a load average around 2 and
ejabberd over 100% CPU usage.

I don't know whether 15 seconds is a reasonable interval: if e.g. each
keystroke in a shared Write touches ejabberd, then 15 seconds seems
long; otherwise perhaps it's very short.

Once I realised that the open files resource limit was killing
ejabberd (which took an embarrassingly long time, not helped by
cryptic log messages), it was stable under all loads.  From time to
time I tried sharing activities between XOs and they were always
responsive.


Douglas