[asterisk-biz] Performance on Pentium Dual Core CPUs

[Matthew Rubenstein]

	These hard capacity numbers are very useful, though not exactly
comprehensive or conclusive. There are too many different factors, some
of which could be nonlinear, to deduce a complete model from which to
extrapolate actual capacity from HW specs.

	However, they do seem to indicate that performace is deterministic:
every type of operation in the callpath consumes its resources
consistently. That consistency suggests that a real benchmark for the
minimum callpath, and benchmarks for "extras" like recording to HD (or
RAMdisk), transcoding, extra legs in MeetMe (or different conference
modules), are available to a lab which just does the typical scientific
testing. As well as testing the same SW and demand on different HW, like
Intel vs AMD, multiple cores, different buses, etc. Those numbers should
be pretty reliable, once established.

	Since Digium will release HW accelerators soon, we will likely start
seeing such benchmarks. But of course those will be as reliable as any
other vendor-only benchmarks in the computer or telephony industry: not
very. Even benchmarks on open Asterisk SW and fully specified HW will
get a chance to cherrypick, unless there's independent benchmarks to
compare them to. If you can produce those benchmarks first, you'll help
put the entire picture on a scientific basis. And maybe get Asterisk
into the pages of a review publisher, like Tom's Hardware, HardOCP, or
even PC Magazine.


On Sun, 2006-06-04 at 14:01 -0400, Sergey Kuznetsov wrote:
> Boris,
> 
> Thanks a lot!
> You proved my thoughts.
> 
> 
> All the Best!
> Sergey.
> 
> 
> Boris Bakchiev wrote:
> > Kevin,
> >
> > Just to prove the point that Asterisk is great I got 2 SQL servers
> > (couple of 6x cpu xeon servers) to host as sip peers.
> > Those are much faster then the previous Celerons.
> >
> > The results kinda puzzled me.
> > With 300 calls I get:
> > Cpu0  : 21.9% us, 10.0% sy,  0.0% ni, 60.2% id,  0.0% wa,  0.0% hi,
> > 8.0% si
> > Cpu1  :  0.0% us,  1.0% sy,  0.0% ni, 99.0% id,  0.0% wa,  0.0% hi,
> > 0.0% si
> >
> > Just got a friend in US who works for ITSP there to test the same
> > scenario with Digium codec on almost identical server (same CPU, mem,
> > motherboard)
> > Cpu0  : 16.4% us
> > Cpu1  :  0.5% us
> >
> > So it appears that Digium codec is better performing then IPP based one.
> >
> > Show channels shows:
> > 600 active channels
> > 300 active calls
> >
> > Sip show channels shows 300 alaw and 300 g729 channels.
> >
> > Iptraf shows:
> > Incoming rates:    3510.0 kbits/sec
> > 6200.2 packets/sec
> > Outgoing rates:    1284.1 kbits/sec
> > 			 1114.4 packets/sec
> >
> > And I can see a flurry of the traffic on the interfaces.
> >
> > As I'm writing, SMP rebalanced the load so it is now:
> > Cpu0  : 10.9% us, 11.9% sy,  0.0% ni, 71.1% id,  0.0% wa,  0.0% hi,
> > 6.0% si
> > Cpu1  :  8.4% us,  4.9% sy,  0.0% ni, 86.7% id,  0.0% wa,  0.0% hi,
> > 0.0% si
> >
> > Are my missing something?
> > Perhaps the conditions are "near perfect" but the encoder and decoder
> > still doing they job and on take 10% of the systems capacity.
> >
> > I think the whole test was affected by the CPU's cache. I think all the
> > data was cached and since server was doing nothing else but transcoding
> > all active code and data was cached in CPU's L2.
> >
> > This makes all kind of benchmarking useless as far as I can see as it is
> > almost impossible to replicate a working system with test tools.
> >
> > I welcome the idea of hardware G729 as it would be a more sensible
> > approach for production systems.
> >
> > Regards
> >
> >
> > -----Original Message-----
> > From: asterisk-biz-bounces at lists.digium.com
> > [mailto:asterisk-biz-bounces at lists.digium.com] On Behalf Of Kevin P.
> > Fleming
> > Sent: Monday, 5 June 2006 02:53
> > To: Commercial and Business-Oriented Asterisk Discussion
> > Subject: Re: [asterisk-biz] Performance on Pentium Dual Core CPUs
> >
> >
> > ----- Sergey Kuznetsov <asterisk_biz at deeptown.org> wrote:
> >
> >   
> >> Can you share your expectations how many G729 transcodings can be done
> >>
> >> on dual dual-core Opterons or P4 with 4 Gb of memory on SIP-to-SIP 
> >> environments ( no TDM or echocancelers involved)
> >> My guess is about 250-300 simultaneous calls.
> >>     
> >
> > I don't have any first-hand experience, as we don't do performance
> > benchmarking like that.
> >
> > In any case, the amount of memory will make no difference at all. The
> > clock speed of the CPUs and the speed of the memory bus will be the
> > determining factors, so realistically I would expect the Opterons to
> > perform better than P4s due to their much more efficient memory
> > subsystem.
> >
> > However, I can say that I doubt any existing PC-type platform will be
> > able to achieve 250 or 300 simultaneous G.729 transcodes; any single box
> > that could handle that would cost far more than a pair of
> > lower-performance systems to do the same thing (think scaling
> > horizontally, not vertically). In addition, when Digium releases the
> > hardware G.729 transcoder board in the near future this will become less
> > of an issue for deployments of that size.
> >
> >   
> 
> _______________________________________________
> --Bandwidth and Colocation provided by Easynews.com --
> 
> asterisk-biz mailing list
> To UNSUBSCRIBE or update options visit:
>    http://lists.digium.com/mailman/listinfo/asterisk-biz
-- 

(C) Matthew Rubenstein