[VoiceOps] Hardware transcoding solutions?

[Simon Perreault]

Le 2015-08-27 17:00, Moises Silva a écrit :
> I'm curious. Can you be more precise on what you mean exactly by generic
> 1U server? What matters mostly is the processor, not the rack unit, so
> I'd be interested in hearing your experience with transcoding 5000 g729
> to g711 channels.

I'm sorry, I can't, for obvious reasons.

> Here are some precise numbers that I run on a dual xeon 1U server
> earlier this month for a presentation at the FreeSWITCH conference, ClueCon.
> 
> The tests were run using 1,000 calls at 50cps (using FreeSWITCH).
> However, I did test at lower cps, same max calls and results are quite
> similar because the bulk of the load is the transcoding

That is my experience as well.

> as it can be seen in the graphs below.
> 
> Without transcoding (just PCMU to PCMU bridging with standard 20ms
> packetization):
> 
> http://moythreads.com/vmstat/xeon-sqlitefullmem-lwarn-sprof-glibc-1000@50cps/report.html
> 
> And with G.722 to PCMU transcoding:
> 
> http://www.moythreads.com/vmstat/xeon-sqlitefullmem-lwarn-sprof-glibc-transcodingG722-1000@50cps/report.html
> 
> At 1,000 calls you already maxed the CPUs with G.722 transcoding, see
> the last graph, that's what's more relevant, showing how the idle cpu on
> the server went from 80% to 15% idle on average just when adding
> transcoding. Granted, there's more powerful boxes out there, but this
> should give people a clear idea of the cost of transcoding.

...on FreeSWITCH. I'd be curious to see the output of "perf top", to see
if CPU time is being spent wisely.

> The algorithmic complexity of G722 is on par with G.729, in fact, G.729 is a
> bit more complex.

That is my experience as well.

Simon