On 3/6/06, Andrew Kohlsmith <[EMAIL PROTECTED]> wrote: > On Friday 03 March 2006 12:44, Shidan wrote: > > Echo can. and trancoding (voice not protocols) are inherently costly, > > and software is still not the best place for them, thats why good > > cards still matter (thats not really why good cards matter in * > > because neither Digiums or Sangomas do this yet and for the cards that > > do, none of those things have been interfaced with zaptel yet). But > > good cards still do matter with *!! > > Echo can is best done as close to the source of echo as possible; having the > echo can *after* the PCI bus just isn't the best place for it, regardless of > costliness.
Yup you're right, the echo can isn't costly, but hardware is still the best place for it. But for transcoding the actual algo's really benefit from extreme parallelism, like the kind given by a good vhdl stack. > Remember that while Asterisk itself may not be optimized for multiple > processing units, the Linux kernel can punt threads off to multiple > processors and the drivers can make use of having different execution units > as well to do their number crunching on, which is exactly what echo > cancellation is all about. Totally agreed > ANY multithreaded application will benefit from multiple processors if the > operating system can effectively schedule the threads to make use of > concurrent execution and the application isn't written poorly. The problem with * threading is it's clean up thread which is like a 60km road sign on the German autobahn. But for enterprise telephony and traffic that isn't carrier grade it wont be an issue and dual processors will always help with all the child threads. and all the external forking and processes, etc. > > As far as whats needed to make * work effectively, a sangoma card on > > a P4 can handle 200 calls/second, with no transcoding. > > Those are really short calls! :-) > I wish the support calls I get were that short!
