On 2012 Feb 16, Thu 14:17:38 you wrote:
> o Robert Szokovacs on 02/16/2012 02:01 PM:
> > Hi,
> > 
> > While tuning our SEMS-based application, we noticed that receiving RTP
> > significantly lowers the number of concurrent sessions with acceptable
> > audio quality (we achieved around 1800 concurrent sessions on a 4-core
> > machine).
> that seems to be pretty few. Are all cores then fully loaded, or do
> you experience higher jitter or delay already earlier (i.e. lock
> contention; CPU still at low usage, but queues filling up etc).

We used 3 media threads, and the load was around 300%. The 1800 was with low 
jitter. We did not really push higher, because we were testing with the 
default MAX_RTP_SESSIONS=2048. (Any particular reason for that value?)

> 
> what codec are you using? how many media processing threads?

PCMA codec.

> You may also check out the multi RTP receiver threads introduced more
> recently, and there were more performance related changes more
> recently (your patch looks like to some 1.4 version).

We will test those once 1.5 is out :)

[...]

> > Is there any drawbacks to this approach?
> 
> I don't see any, if you don't need the audio, or in-band DTMF. The
> DTMF event queue of the session is protected by its mutex, so it's
> safe to post events there from the RTP receiver threads the same way
> as the media processing thread. You have your sessions in the
> MediaProcessor, I guess, so DTMF should be processed and passed to the
> applicaiton.

So, shall I submit a proposed patch then?

br

Szo
_______________________________________________
Semsdev mailing list
[email protected]
http://lists.iptel.org/mailman/listinfo/semsdev

Reply via email to