[Asterisk-Users] Echo canceller data-points

[Steve Davies]

On 3/16/06, Andrew Kohlsmith <akohlsmith-asterisk at benshaw.com> wrote:
> On Wednesday 15 March 2006 16:46, Steve Davies wrote:
> > I thought I would try the 1.2 trunk/HEAD version of MG2 with the extra
> > knobs and whistles, but found 2 problems. This version trains even a
> > normally clean line in about 10 seconds, rather than instantly, and
> > its CPU usage is through the roof compared to the 1.2.4 version of the
> > code. (FYI I got very similar resuls at all intermediate SVN versions
> > of the MG2 canceller between 1.2.4 and HEAD)
> >
> > My advice: Go with the 1.2.4 MG2 echo canceller, perhaps if you have
> > plenty of spare CPU the newer code will become useful, but I could not
> > cancel even 20 PRI channels using a 1GHz processor on the latest code
> > - I got clicks, buzzing and eventually a dead PRI. With the 1.2.4
> > branch I had 40% CPU free when cancelling 30 channels.
>
> What did you use for CFLAGS/KFLAGS and PROC for your zaptel build?  What about
> USE_MMX and CALC_XLAW in zconfig.h?
>

I leave both of these disabled in both builds - CALC_XLAW says it
optimises for "small" number of channels without defining "small", and
MMX has warnings about not using it on AMD processors - We don't use
AMD, but we use non-INTEL (Via) chips, so we play it safe. CFLAGS and
KFLAGS are also the defaults of -O4 and -O6 respectively.

My real concern is the apparent CPU usage delta between MG2 in 1.2.4
and MG2 in trunk, but even when only cancelling one channel, the trunk
version seems less effective and trains much more slowly.

On the plus side, we now have 4 sites running the 1.2.4/MG2 code, and
they have all seen benefits from it.

Regards,
Steve