[asterisk-biz] Recent REMOTE CRASH BUG

Trixter aka Bret McDanel trixter at 0xdecafbad.com
Sun Mar 4 18:47:43 MST 2007


On 3/4/07, Kevin P. Fleming <kpfleming at digium.com> wrote:
>
> Trixter aka Bret McDanel wrote:
> > instead of much more frequently.  I would also like to see interrupts
> > once per span (possibly per card) as opposed to for each channel.  These
>
> This is how it already works. One interrupt per span per millisecond. If
> interrupts were generated for each channel the system would fall apart
> rapidly with any significant number of spans in it.


for part of what I said, ok I was misinformed.  However 1 per ms is still
10-20 times more frequently than is generally needed (there are exceptions
but they are generally rare).

I also dont think that a DMA ring buffer exists (*please* correct me if I am
wrong) so that it doesnt take the cpu to blit the data off the card into
ram.  That isnt that hard to do, you just write to the bus address specified
in the card init, which is a value reserved by the kernel.  The card needs
to be able to write to those addresses, which since its on the bus it has
access it just needs the firmware to allow it.

You could fire an interrupt once the buffer is full (for those that dont
know in a ring buffer you have multiple buffers that are done like a hunt
group, first, second, third, first, ...) which could be say 10ms.  This also
makes packetization easier, it would generate a lot fewer context switches
and generally make a better product capable of handling more on the same
box.

Zaptel would have to be updated since it doesnt really support that method,
instead relying on the 1000 interrupts per second thing instead of 50 per
span.

and for those that would say the memory requirements would be too large,
10ms of g.711 audio is roughly 6400 bytes per DS0 which would be 150kB per
T1 (less if pri, more if E1).  In a world of gigs of ram that is a trivial
price to pay for the performance gains that would be realized.  Even on a
ram limited system, the performance gains would likely be higher since they
also generally have limited cpu resources to handle all those context
switches.


-- 
Trixter http://www.0xdecafbad.com     Bret McDanel
Belfast +44 28 9099 6461        US +1 516 687 5200
http://www.trxtel.com the VoIP provider that pays you!
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.digium.com/pipermail/asterisk-biz/attachments/20070304/22a18eab/attachment.htm


More information about the asterisk-biz mailing list