[Asterisk-Users] Sangoma E1 board Experience

[janvb at caselaboratories.com]

>You're still not answering his question.  :-)
>
>The TJ320 has no buffering capability but the Silabs parts have nothing to do 
>with that, and neither do the framers on the higher-density TDM cards.  I 
>think his question is more "why don't these cards have bigger PCM buffers and 
>interrupt less often, or at least have deeper buffers so if an interrupt is 
>delayed I don't get overruns?"
>
>I believe the answer lies in latency.  You do *not* want deeper buffers.  
>Better interrupt handlers, perhaps, but not any deeper buffering on the 
>hardware, as that just increases latency.
>  
>
This is a tradeoff. You would ideally want a 0.125 ms latency which 
actually is possible (not on these boards thought), but since you need 
to transport through a PCI and use a host you need to adjust to a 
reasonable tradeoff. I would prefer to use 6ms (48 bytes), which is a 
more common handrule for PABX design and a far more reasonable interrupt 
speed for a Pentium.

You will hit 2 perfomance issues with a 1ms packet.

- No pentium are particulary efficient moving packets below 50 bytes. 
Make a small C program that move from 1 to 100 bytes and you will see 
that the xfer capacity increase dramatically as you increase packet 
size. This don't really matter if you use DMA, but it matters a lot if 
the Pentium is doing the job.

- Some of the older type pentiums which you typically find in embedded 
boxes are very inefficient handling fast interrupts and might use 10-20% 
of their capacity just serving the interrupt in this case. This drops 
dramatically if you are able to increase to 6 or 10 ms packets, which 
also is a reasonable tradeoff as far as latency goes.

Also remember that echo cancel, dtmf, G.729.A etc all need more that 1ms 
packets to work on, meaning you have to buffer anyway. dtmf can happen 
outside the voicepath, but echo cancel and g.729.a has to be in your 
voicepath.

But, thanks for your answers - I realize there is little that can be 
done to address this issue as none of these boards seems to have dynamic 
buffering.

jan