[OpenSIPS-Users] Setting opensips for maximum performance

[Bogdan-Andrei Iancu]

Hi Jan,

What Stan says is true - also when I'm doing the sipp tests, the biggest 
problem is the sipp itself - it get overloaded faster than opensips. And 
not only the CPU is important here, but especially the I/O - sipp may 
also loose packages (discard) and report them as failures. Also, I'm not 
sure how performant is the virtualization when come to intensive I/O.

So, try also spreading the sipp on more physical servers and reduce the 
load.

Typically, in such tests, the approach is bottom - up : you start with 
small load and keep increasing it until you get problems - then you 
start troubleshooting the problem to see how's responsible for 
it....when fixed, you start increasing the load again.

The test beds are like puzzles and not all the time the "tested" entity 
is the guilty one - also may be the net, the tester, etc.

Regards,
Bogdan

Jan Rozhon wrote:
> Hello Stanislaw,
>
> thanks for your reply. I dont think that trouble causes SIPp. I run all 
> 6 instances on 6 virtual computers - the physical box contains 16GB of 
> RAM and 8 Xeon 2,8GHz cores - for each virtual computer I allocate 1 
> core of physical CPU, since SIPp is just single threaded , and 1024 GB 
> of RAM - OS on VM is Ubuntu 9.10. On a very old machine I was able to 
> reach 200 calls without bigger issues, now on high tech server I am 
> limited to 200 calls per second as well...
>
> The exact problem is, that the call is interrupted before it is ended by 
> regular Bye/200Ok. Until I reach 200 calls per second - everything is 
> fine, then retransmissions rapidly increase and after 240 calls the 
> significant number of calls is lost in progress.
>
> Dont know where to look -
> on PbX
> -CPU is not fully utilized
> -RAM is used just to 25% (around 1024MB)
> -LAN is not overwelmed
> -UDP buffer discard just few datagrams
>
> on VMs with SIPp
> -CPU is utilizet to 50% tops
> -RAM is OK
> -LAN is ok
> -UDP buffer ok too
>
> Thanks, Jan.
>
> Dne 21.2.2010 20:19, Stanisław Pitucha napsal(a):
>   
>> On 18.02.2010 13:57, Jan Rozhon wrote:
>>    
>>     
>>> AMD Athlon processor. As a generator of SIP traffic I use SIPp v3.1
>>> running on 4 virtual computers as UAC and two computers as UAS all
>>>      
>>>       
>> You're not running them all on the same physical box, are you?
>> If yes, then you might find sipp itself to be a CPU hog and causing a
>> lot of trouble + overhead of processing 6 machines + overhead of
>> software networking... You should be able to push 200 SIP-only CPS from
>> a single box / single sipp process without problems - I'm able to do
>> 100+ cps on a dated single-core laptop.
>> Start small - then increase until you see problems and check why exactly
>> did they occur.
>>
>>    
>>     
>>> calls) many of those calls (aroud 50%) don't finish succesfully. CPU
>>>      
>>>       
>> Can you find out what does "don't finish successfully" mean exactly? Are
>> you not getting the same packets on the other side? Are you hitting the
>> retransmission limit? Are you getting wrong packets? It's easier to
>> debug stuff once you know exactly how it fails...
>>
>> Some more stuff to check: are you using retransmissions in sipp? how
>> long does it take to process a single message in opensips (benchmark
>> module helps here a lot)?
>>
>> _______________________________________________
>> Users mailing list
>> Users at lists.opensips.org
>> http://lists.opensips.org/cgi-bin/mailman/listinfo/users
>>    
>>     
>
>
> _______________________________________________
> Users mailing list
> Users at lists.opensips.org
> http://lists.opensips.org/cgi-bin/mailman/listinfo/users
>   


-- 
Bogdan-Andrei Iancu
www.voice-system.ro