[Asterisk-Users] Transcoding CPU usage: surveys?

[John Todd]

Before I put myself to the task (next month, maybe) of surveying the 
CPU costs of transcoding, perhaps someone else has already done this 
work and would be willing to share it or refer me to a link of 
previously published data.  My reviews of the mailing list with 
various keywords were unsuccessful in finding adequate references, 
though I admit I only spent 20 minutes looking.

What I seek is this:

   Using a "standard unit of processing power", such as a 2ghz P4 
processor with a bottleneck-free 100baseT connection between 
endpoints, at what point does voice quality noticeably degrade in the 
following circumstances:

    1) SIP G.711 to IAX2 GSM
    2) SIP G.711 to IAX2 iLBC
    3) SIP G.711 to IAX2 g.729
    4) Zap to IAX2 GSM
    5) Zap to IAX2 iLBC
    6) Zap to IAX2 g.729
    7) Zap to IAX2 g.711
    8) Zap to SIP GSM
    9) Zap to SIP iLBC
   10) Zap to SIP g.729
   11) Zap to SIP g.711
   12) TDMoE to SIP G.711
   13) TDMoE to SIP GSM
   14) TDMoE to SIP iLBC
   15) TDMoE to SIP g.729
   16) TDMoE to IAX2 GSM
   17) TDMoE to IAX2 iLBC
   18) TDMoE to IAX2 g.729
   19) TDMoE to IAX2 g.711

   20) Reverse tests 1-20 to see if there is a difference

I exclude IAX2 to IAX2 transcoding, since I think that SIP to IAX2 
should be close enough to compare.  I also exclude passthrough 
traffic (no transcoding) on all protocols though that would be 
interesting if anyone has it.

I realize that some of these tests may be redundant, and some have 
such minimal differences as to make testing somewhat silly.  However, 
any subset of the data would be great to have in a "known" testing 
environment instead of rumor, so I exhaustively list them for my 
future test construction.

The term "noticeably degrade" is of course a matter of opinion, but 
your opinions are still better than nothing.  :-)  I am very 
interested in anyone who can quote me raw figures for packets-only 
network.  In other words, no Zap or TDMoE channels - just SIP-to-SIP, 
or SIP-to-IAX2, or IAX2-to-IAX2 type of circumstances, as I think 
these are more representative of what the processor (versus the 
backplane and IRQ peculiarities) represent in possible throughput.  I 
honestly don't care what processor or speed you tell me you've used, 
as long as you tell me exactly what the system looked like for your 
experiment(s).

JT