[asterisk-dev] Re: Kernel modules => mainline kernel

[Andrew Kohlsmith]

On Tuesday 13 February 2007 3:00 am, Oron Peled wrote:
> Q: What is the interrupt rate in this situation?
> A: Assume a favorable condition: All PRI with one interrupt
>    per port (with analog lines we would need many more cards).

Let's assume a realistic situation: a quad PRI card with one interrupt per 4 
ports.

>    Number of required cards =~ 1000 / 30 = 34 (assume we have such a PC ;-)
>    Number of interrupts = 34 * 1000/sec = 34,000/sec (without counting
>    the matching context switches on the transmit path).

1000 / (30 * 4) = 8.3 = 9000 interrupts/sec.

If we can use NFAS on each quad card, that gets us 8.1 cards, and if 
Zaptel/libpri allows us to NFAS multiple cards (I don't see why it wouldn't), 
we would have 8 cards, not 9.

I've long, long been wondering if it's possible to slave additional Zaptel 
cards off of one card, but I haven't the real need for it to go about 
actually implementing a test.  That would certainly eliminate some of your 
arguments here.

> For anyone interested, during kernel 2.2 everybody were talking about
> zero-copy. Then came the reality check. Zero copy only helped Gigabit
> ethernet and above. However, the NAPI api that enabled drivers to
> dynamically switch to polling instead of interrupt per-packet was a
> major win for many 100Mb drivers (the results were applied to kernel 2.4).

Yes, and it shot latency right to hell by saving all these interrupts.  While 
it might be a solution if we used the high-resolution timers in the later 
kernels, I don't think it would be feasible any other way.

-A.