[asterisk-users] Ast/Hyla/IAX Scalability?

[Steve Underwood]

David Backeberg wrote:
> On Fri, Mar 13, 2009 at 2:30 PM, Marshall Henderson
> <marshallmch at gmail.com> wrote:
>   
>> On Fri, Mar 13, 2009 at 11:07 AM, David Backeberg <dbackeberg at gmail.com> wrote:
>>     
>>> Again, you'll find people arguing that their voip solution has as low
>>> of a failure rate as a hardware solution. I'm jealous. My voip fax
>>> solution does not yet have that low of a failure rate, but I'm
>>> hopefully getting closer to working out the last bugs.
>>>
>>>       
>> Do you have any specifics to share about the problems you're finding?
>>     
>
> Sure. I can't disagree with the poster who said that problems they've
> seen are really the other side's fault. But assigning blame doesn't
> make me any happier. I have fax receiving problems I can't reproduce.
> When I load test it, I don't have problems. When I send 'real' inbound
> faxes from outside the network, over the real phone system, I don't
> have problems.
>
> I'm in New Haven, CT. One sender that messes up the most is in Kansas
> City, KS. They are a legitimate client, really sending a fax. I get
> occasional fax receipts that say:
>
> 'The call dropped prematurely'
>
> There will sometimes be a cluster of these, followed by a successful receipt.
>
> When I load tested, and send from real fax machines out and back in on
> POTS, I get 100% success. I've successfully load-tested around 175
> simultaneous inbound faxes. I slowed down the simulation to about 5
> simultaneous faxes, and left that running over a long weekend,
> generating something like 30,000 faxes and something like 1GB of
> received fax files. Again, the success rate was 100%. A problem with
> my simulation was that I used sending faxes that speak the protocol
> correctly. Does anybody have some faxes that send garbage?
>
> Then I put it into production with a limited amount of real fax
> traffic for our clients. I'm talking fewer than 10 calls per day most
> days. But it seems like the reality of the speed of light over
> continental long-distance, combined with the reality of crappy fax
> machines that don't speak protocols correctly result in occasional
> failures. I've made some adjustments that I think anecdotally have
> solved the silly problems, but that one with the faxes dropping early
> is the one that (maybe) hasn't gone away.
>
> I'd like a success rate around 99%. I'm getting around 63% if you
> count individual failed calls that eventually result in a success. I
> can't tell if I'm having bad luck with this phase of my pilot or if my
> failure rate is going to remain constant as I add clients. I need more
> data points to get statistical significance. What I really need is a
> failing fax I can control, then tune parameters on my side, and see if
> the failure rate gets worse or better. Seriously considering breaking
> down and asking for the cooperation of the client in that endeavor.
>
> People who have been following my posts on this topic know that I'm using:
> PRI(s) -> Cisco voip gateway hardware -> T.38 / SIP / g711 ->
> Asterisk-1.6 with ReceiveFax (depends on SpanDSP, but does NOT use IAX
> or IAXmodem)
>
> What I've been 'tuning' most recently have been arguments to the Cisco
> setup fax and SIP translation.
>
> I did try out IAXModem with Hylafax and 1.4 and had lots of problems
> that all went away when I switched to using the approach I use now. I
> never tried 1.6 with IAXModem and Hylafax, so I can't tell you how
> well they work together.
>   
Fully open-to-the-public FAX servers tend to get just get a lot of bad 
calls, many of them wrong numbers, or voice users. FAX servers for 
closed user groups tend to get few bad calls, unless the phone number 
gets included on some unfortunate list. This is one of the things which 
made early real world testing of spandsp and iaxmodem tough. We have to 
capture every failure, and analyse them by hand whether it was our fault 
or the far end's. Without knowing the nature of your system I have no 
clue what kind of failure rate might be expected. You can find a bit 
more about these issues and our results at 
http://www.soft-switch.org/spandsp-soft-fax-performance.html

Your differing failure rates between using ReceiveFAX and using iaxmodem 
seem to indicate your results relate to issues in your own system, 
rather than the nature of the callers, but we can't really tell. A minor 
change in usage pattern may have resulted in a big change in the 
results. What I can say is that a properly set up iaxmodem + HylaFAX 
setup, with an IAX connection that does not loose packets (don't assume 
LANs don't loose packets), will have a true failure rate (i.e. a rate of 
calls failing which had the potential to succeed) well below 1%. The 
results for the latest spandsp used on its own (i.e. as a full FAX 
machine, rather than a FAX modem for HylaFAX) should be approaching this 
figure, as its back end processing has matured. Right now I think the 
maturity of HylaFAXes handling of buggy FAX machine probably still puts 
it slightly ahead. Success with FAX has a lot to do with tolerating the 
quirks of a 1000 buggy designs, and it takes time for this to shake out.

As for the volume of FAXes a machine can handle, the question is like 
"how long is a piece of string?". No two people do quite the same thing. 
Sending and receiving cause different processor loads - sending FAXes is 
generally more lightweight than receiving them. Doing front end TIFF 
generation or, say, back end TIFF to PDF conversion adds load, which 
needs to be considered in the total machine load. For simple FAXing many 
modern machines can cope with hundreds of channels. Your results may vary.

Regards,
Steve