[Freeswitch-users] mod_conference scalability

[Steve Underwood]

On 12/18/2009 03:29 AM, Brian wrote:
>
> Hi Mike,
>
> I didn’t get around to testing on the FreeSWITCH trunk yet. Are there 
> substantial fixes to mod_conference in the FreeSWITCH trunk that might 
> increase capacity for my scenario of one speaker and many listeners? 
> If I want to put this into a production environment, I would need a 
> stable version, which as far as I know is the 1.0.4 version.
>
> However, I did test on Asterisk 1.4 using app_conference, and doing 
> the same scenario was able to get 1 speaker and 600 listeners on a 
> single conference with no audio issues. The CPU at that point was just 
> over 300%, same as where the single conference scenario failed on 
> FreeSWITCH with 300 listeners. I was able to push it to over 700 
> listeners before I reached 400% CPU usage (I guess maxing out my 
> quad-core processors), and asterisk finally crashed. But up until that 
> point, there were no audio problems.
>
> I’ve read a lot about how FreeSWITCH is supposed to be more scalable 
> than Asterisk, but unless there is something wrong with my FreeSWITCH 
> setup, Asterisk was clearly the winner in this test – more than 
> doubling FreeSWITCH capacity in this case. Again, maybe there is 
> something on the FreeSWITCH side that I’m doing wrong, but I don’t see 
> what it could be.
>
> Brian.
>

I don't think you have mentioned which codecs are involved. This can 
have a profound effect.

Steve

> *From:* Michael Jerris [mailto:mike at jerris.com]
> *Sent:* Thursday, December 17, 2009 10:18 AM
> *To:* freeswitch-users at lists.freeswitch.org
> *Subject:* Re: [Freeswitch-users] mod_conference scalability
>
> I would be curious what the same tests produce with svn trunk of 
> FreeSWITCH.
>
> Mike
>
> On Dec 16, 2009, at 4:49 PM, Brian wrote:
>
>
>
> Hi,
>
> I’m new to FreeSWITCH and I’m testing the scalability of 
> mod_conference to see if it will scale better that other solutions. My 
> scenario is to have one speaker, and many listeners (mute). Since I 
> have only one speaker, I was expecting this to scale well because 
> there is no audio mixing required, just send each frame of the single 
> speaker to each listener. Unfortunately, my testing was disappointing, 
> and it didn’t scale nearly as well as I’d hoped (based on what I’ve 
> read on how FreeSWITCH is supposed to be generally very scalable).
>
> Here’s my server setup is this:
>
> FreeSWITCH 1.0.4, 64 bit CentOS 5.3, on a quad-core Xeon server, 4 Gig 
> of RAM. I’ve set file logging to “notice” level. My conference profile 
> is configured to suppress several events, hoping that it would improve 
> performance.
>
> Here are a few scenarios I tested, and roughly where I reached the 
> point of audio failure on the conferences:
>
> Scenario 1:
>
> 1 conference, 1 speaker, audio failed at approx 300 listeners (mute)
>
> Scenario 2:
>
> 4 conferences, 1 speaker per conference, audio failed approx 110 
> listeners per conference (so just over 400 total channels on the system).
>
> Scenario 3:
>
> 16 conferences, 1 speaker per conference, audio failed at 32 listeners 
> per conference (so just over 500 total channels on the system).
>
> Looking at the output from “top”, it seems that in all 3 scenarios, 
> the audio quality failed when the % CPU for the FreeSWITCH process 
> exceeded 300%.
>
> I was hoping maybe someone else might have done similar testing, or 
> maybe has suggestions on how to improve the performance. Or perhaps an 
> alternate solution to the one speaker, many listener case?
>
> Thanks,
>
> Brian.
>
>