Hi dear friends, We tried to benchmarking IPtel SER 0.9.6 on Intel xeon CPU based computer recently,but we met a problem about poort performance.For example on a computer configured with Intel 2*sossaman CPU (four cores)+2G ddr2-400 memory+100M Ethernet card, we didn't configure any database,and registered with only 1-2 users,and we use SIPP 3.0 as the call generator ,but we can only see about 3,000 calls per second for a very simply built-in scenario in SIPP(UAC+UAS),and the CPU idle percentage is about 70-80%.if we tried to increase the call rate ,we got failure calls.
We tried to enlarge the memory to 768M for SER,and limited the ser processes number on each port from 8 to 4,and we also made some improvement on config.h & Makefile.defs,but didn't get big improvement on performance . If anybody could help me on this,I will appreciate a lot.
Thanks Fang tian.
Hi Fang,
Do you consider it realistic to produce load for millions of users with just 1-2? I don't. What probably happens is that you get a lot of inter-locking somewhere like in the registrar because you use just 2 users. So you could have 100 CPUs and gigs of RAM, but if you do not use more users, it won't scale.
I would say that you need to use ratios for calls/sec/users way smaller than 1, while you now seem to be at about 1500 or so... This is highly unrealistic and also not relevant.
Cheers, -Dragos
fang, tian wrote:
Hi dear friends, We tried to benchmarking IPtel SER 0.9.6 on Intel xeon CPU based computer recently,but we met a problem about poort performance.For example ,on a computer configured with Intel 2*sossaman CPU (four cores)+2G ddr2-400 memory+100M Ethernet card, we didn't configure any database,and registered with only 1-2 users,and we use SIPP 3.0 as the call generator ,but we can only see about 3,000 calls per second for a very simply built-in scenario in SIPP(UAC+UAS),and the CPU idle percentage is about 70-80%.if we tried to increase the call rate ,we got failure calls. We tried to enlarge the memory to 768M for SER,and limited the ser processes number on each port from 8 to 4,and we also made some improvement on config.h & Makefile.defs,but didn't get big improvement on performance . If anybody could help me on this,I will appreciate a lot. Thanks Fang tian.
FREE Animations for your email - By IncrediMail! Click Here!
http://www.incredimail.com/index.asp?id=105578&rui=93057459
Serusers mailing list Serusers@lists.iptel.org http://lists.iptel.org/mailman/listinfo/serusers
Dragos Vingarzan Dragos.Vingarzan@fokus.fraunhofer.de wrote:
Do you consider it realistic to produce load for millions of users with just 1-2? I don't. What probably happens is that you get a lot of inter-locking somewhere like in the registrar because you use just 2 users. So you could have 100 CPUs and gigs of RAM, but if you do not use more users, it won't scale.
I would say that you need to use ratios for calls/sec/users way smaller than 1, while you now seem to be at about 1500 or so... This is highly unrealistic and also not relevant.
This _is_ realistic in case SER is used as proxy on a path between e.g. vendor gateways. In that case it can have 0 (i.e. _zero_) "users" but thousands of calls per second. And, of course, this shall be relevant unless you're positioning it to "office PBX" which is definitely not direct SER niche.
Valentin Nechayev wrote:
Dragos Vingarzan Dragos.Vingarzan@fokus.fraunhofer.de wrote:
Do you consider it realistic to produce load for millions of users with just 1-2? I don't. What probably happens is that you get a lot of inter-locking somewhere like in the registrar because you use just 2 users. So you could have 100 CPUs and gigs of RAM, but if you do not use more users, it won't scale.
I would say that you need to use ratios for calls/sec/users way smaller than 1, while you now seem to be at about 1500 or so... This is highly unrealistic and also not relevant.
This _is_ realistic in case SER is used as proxy on a path between e.g. vendor gateways. In that case it can have 0 (i.e. _zero_) "users" but thousands of calls per second.
May be, but if you will have no users and you will only forward calls from trusted sources, then there will be no database lookups for registers and deregisters as in sipp uas/uac scenario.
Libor
And, of course, this shall be relevant unless you're positioning it to "office PBX" which is definitely not direct SER niche.
That is something else. But Fang said "registered with only 1-2 users" and not just proxy something. So I understood it like Alice, Bob registered and then making the same Alice calls Bob thing 3000 times/second, which will probably end-up as being locked on Bob's lookup in the registrar.
My point is: most of the time you have to generate realistic traffic in order to see real performance. It is not realistic that Bob gets 3000 cps terminated to him... having a registrar that would scale properly with the number of CPUs is overkill especially as this is not realistic.
-Dragos
Valentin Nechayev wrote:
Dragos Vingarzan Dragos.Vingarzan@fokus.fraunhofer.de wrote:
Do you consider it realistic to produce load for millions of users with just 1-2? I don't. What probably happens is that you get a lot of inter-locking somewhere like in the registrar because you use just 2 users. So you could have 100 CPUs and gigs of RAM, but if you do not use more users, it won't scale.
I would say that you need to use ratios for calls/sec/users way smaller than 1, while you now seem to be at about 1500 or so... This is highly unrealistic and also not relevant.
This _is_ realistic in case SER is used as proxy on a path between e.g. vendor gateways. In that case it can have 0 (i.e. _zero_) "users" but thousands of calls per second. And, of course, this shall be relevant unless you're positioning it to "office PBX" which is definitely not direct SER niche.
-
Dragos Vingarzan -
fang, tian -
Libor Chocholaty -
Valentin Nechayev