Tuesday 2 October 2007

Results of stressing ejabberd+mod_multicast with Jabsimul

I tested mod_multicast in a living server, stressing it with synthetically generated load using Jabsimul.

The setup was: create 300 accounts with Testsuite's userreg; create a Shared Roster Group with @all@; configure Jabsimul to login in all the 300 accounts and change presence every 60 seconds.

Note that each user has 299 contacts online, and consequently each presence change generates a presence packet with 299 destination addresses.

I ran Jabsimul against an unaltered ejabberd trunk, and also a mod_multicast enabled version. The CPU consumption in both cases varied around 20% to 30%. I couldn't find a clear difference between enabling or disabling mod_multicast. The Virtual memory was around 125MB, and Resident memory around 105MB.

It seems the computation resources required by mod_multicast are only a small part of all the processing that takes place in ejabberd. So, this test indicates that probably enabling XEP33 in an ejabberd server does not have an appreciable impact in the server CPU or RAM consumption.

The computer used in the tests:

  • AMD Athlon(tm) 64 Processor 3000+, 4.000 Bogomips
  • 1 GB RAM
  • Debian sid
  • Linux 2.6.22-2-686 (Debian package)
  • Erlang R11B-5 (Debian package)
  • ejabberd SVN r952
  • mod_multicast SVN r394

If you run your own tests and find a differing conclusion, please tell me.

PD: Some instructions to get Jabber Test Suite and Jabsimul: Benchmarking Jabber/XMPP Servers with Jabsimul

No comments: