[Asterisk-Users] Re: TDM users: modified zttest.c for testing

[Mike Mueller]

On Tue, May 03, 2005 at 09:28:23PM +0000, Tony Mountifield wrote:
> I wrote:
> > In article <Chameleon.1115133799.adar0 at vegas>,
> > Rich Adamson <radamson at routers.com> wrote:
> > > 
> > > It would be very interesting to see everyone's results in running
> > > this, and even more interesting to report the results with the OS
> > > distro in use, mobo in use (if known), etc. If anyone actually
> > > get's a result that is very close to 1.000 seconds, I'd really
> > > like to know more about those systems. (email off list is fine
> > > if you want.)
> > 
> > --- Results after 20 passes ---
> > Best: 1.024003 -- Worst: 1.023981 -- Average: 1.023993
> > 
> > This looks very close to 1024ms instead of 1000ms. That got me thinking:
> > 
> > I believe your premise is wrong. The sample rate of telephony audio
> > is 8kHz. With 8-bit samples (uLaw or aLaw), that means 8000 bytes
> > should be supplied in 1 second, not 8192.
> > 
> > At a rate of 8000 bytes/sec, 8192 bytes will arrive in 1.024 seconds.
> 
<snip> 
> [root at softins zaptel]# ./zttest-mod -v
> Objective: to read 8000 bytes from TDM card in 1.000000 seconds.
> Opened pseudo zap interface, measuring accuracy...
> 
> read(fd, buf, 8000) returns 1024
> read(fd, buf, 6976) returns 1024
> read(fd, buf, 5952) returns 1024
> read(fd, buf, 4928) returns 1024
> read(fd, buf, 3904) returns 1024
> read(fd, buf, 2880) returns 1024
> read(fd, buf, 1856) returns 1024
> read(fd, buf, 832) returns 832

Whew! At least the kernel module is using the len :).

> 8000 bytes in 1.023988 seconds
> read(fd, buf, 8000) returns 1024
> read(fd, buf, 6976) returns 1024
> read(fd, buf, 5952) returns 1024
> read(fd, buf, 4928) returns 1024
> read(fd, buf, 3904) returns 1024
> read(fd, buf, 2880) returns 1024
> read(fd, buf, 1856) returns 1024
> read(fd, buf, 832) returns 832
> 8000 bytes in 1.023998 seconds
> --- Results after 2 passes ---
> Best: 1.023998 -- Worst: 1.023988 -- Average: 1.023993
> [root at softins zaptel]# 
> 
> So it looks like the pseudo driver is always handling 1024 byte chunks,
> and even if you ask it for fewer bytes, it takes 1024 bytes' worth of
> time.
> 
> I think it should really be handling 1000-byte chunks in 125ms rather
> than 1024-byte chunks in 128ms, if it is supposed to be emulating
> telephony channels.

Why? Computers are base-2 oriented and POTS digital telephony is based
on adapting to human hearing perception and a massive installed base of
analog equipment.  Working with base-2 numbers in computer programs is
common and often efficient.

1/8000 = 0.000125 sec/sample octet
(8 * 1024)samples * 0.000125 sec/sample = 1.024 sec

Looks good to me.  

> 
> But zaptel.c is Deep Magic, and I'd be interested in comments from those
> who are famliar with it in detail.

Bah. Just your average 6459 line kernel module ;).  I've seen bigger.
Here are some guides:
http://www.oreilly.com/catalog/linuxdrive2/
http://kernelnewbies.org/documents/kdoc/kernel-api/linuxkernelapi.html

-- 
Mike