Ejabberd resource tests/try 5

Try 5: a few hundred users, interacting properly

It turns out that in the hyperactivity agents were not interacting in Ejabberd resource tests/try 4, due to the shared roster being off by mistake.

For this test several hyperactivity instances were used, each running 50 clients. 4 XO laptops were also connected, so the total number of clients was at most stages 50n + 4.

Memory use

Here you can see the memory use is slightly over 1 MB per user, and growth is approximately linear or slightly in excess thereof. The server has 1GB of ram.

Memory use vs connections

This shows max-median-min values for stable points along the way (3 minutes with the same number of connections).

Here's the raw data (duplicates exist because a number of connections can vary slightly as clients disconnect and reconnect).

resident_mem
clients              minimum     median    maximum
54                        96         97         98
104                      134        139        143
154                      164        167        174
204                      220        224        253
204                      229        232        236
254                      270        278        289
304                      345        399        421
254                      329        348        386
304                      337        352        419
354                      455        471        515
403                      499        577        654
404                      502        504        514
454                      635        644        696
453                      642        644        646
454                      608        613        664

CPU usage

Cumulative CPU use by the ejabberd process:

Load averages, by connections and by time:

The laptops

Note that they all see a different number of clients.