[Server-devel] testing ejabberd

[Gary C Martin]

Hi Douglas,

On 4 Oct 2008, at 05:30, Douglas Bagnall wrote:

> The results are summarised here:
>
> http://wiki.laptop.org/go/ 
> Ejabberd_resource_tests#Try_4:_a_few_thousand_users
>
> In short, with 1GB ram, ejabberd coped with a stable load of 2000
> connections, but it went crazy when faced with more, bouncing off the
> RAM ceiling, dropping clients, and freezing its web admin interface.
> Then after a quiet period it recuperated and gamely made the fatal
> number of connections.

Hey, nice set of graphs.

So, my take on looking through them was that you are, most definitely,  
memory bound. Once you get to ~1500 active users you're out of  
physical memory, IO waits go up, CPU load goes down, VM usage spikes,  
things that really shouldn't start getting swapped to disk, services  
start grinding to a halt left, right, and centre...

I had a colleague who's answer (in part) to pretty well any server  
load issue that came up was "add RAM". He was right, most of the time,  
and adding ram never hurt anyone :-)

One extra figure that would be interesting is the server response  
latency to client requests, not sure if hyperactivity gives you that  
easily. Adding RAM should help to stabilise the system behaviour, but  
you'll start drifting into server responses that are slow enough that  
various layers of the communication stack start to think things have  
broken. At that point there will likely be a spike in retransmissions/ 
reconnections et al, and all that extra traffic will quickly bring the  
server to it's knees again. :-)

All interesting stuff.

--Gary