Performance test results for Mobicents Jain-SLEE 2.0.0.GA

[Luis Barreiro]

Hello all,

Attached are the reports of a performance test that Mobicents Jain-SLEE 2.0.0.GA was subjected to.
10 Million SIP calls were handled by sip11-load test-with-dialogs example at a call rate of 750 calls/second.

The setup for this test includes a 9GB heap with LargePages, as well as some other optimizations both at JVM and Mobicents level that can be found below.

From the SIPp report we can see that there were no call failures.
vmstat report show us that CPU usage was at 85%.


The test was done under the following conditions:

Host: qa03.mw.lab.eng.bos.redhat.com

Processor: Quad-Core AMD Opteron(tm) Processor 2356 (2300 MHz, cache size: 512 KB)

[lbarreir@qa03 ~]$ cat /proc/meminfo
MemTotal:     10235188 kB

[lbarreir@qa03 ~]$ uname -r
2.6.18-136.el5

[lbarreir@qa03 ~]$ $JAVA_HOME/bin/java -version
java version "1.6.0_18"
Java(TM) SE Runtime Environment (build 1.6.0_18-b07)
Java HotSpot(TM) 64-Bit Server VM (build 16.0-b13, mixed mode)

JAVA_OPTS="-Xms9000m -Xmx9000m -Xmn256m -XX:MaxPermSize=128m -XX:+UseLargePages -XX:+UseTLAB -XX:+PrintGC -XX:+UseConcMarkSweepGC -XX:-UseBiasedLocking -XX:+EliminateLocks -Dorg.jboss.resolver.warning=true -Dsun.rmi.dgc.client.gcInterval=3600000 -Dsun.rmi.dgc.server.gcInterval=3600000"

Mobicents Jain-SLEE 2.0.0.GA with 8 threads in event router and MONITOR_UNCOMMITTED_AC_ATTACHS flag set to false.
Sip Resource Adaptor with 8 threads and network buffers set to 128k.
Sip11-loadtest-with-dialogs example from SVN. All logs disabled.

Load generator: SIPp 3.1, using the scenario uac.xml that can be found in SVN along with the example, started with the command:
./sipp 127.0.0.1:5060 -sf uac.xml -i 127.0.0.1 -p 5050 -r 750 -m 10000000 -l 1000000 -nd -fd 1 -trace_stat

--
Luis Barreiro
Mobicents Quality Engineering
JBoss^® by Red Hat

[aayush bhatnagar]

Amazing. Congrats. Very nice figures..

--
aayush
http://in.linkedin.com/in/abhatnagar19
======

[Julian]

Hi,

What happened to performances test results? aren't they public anymore?

Best regards.

[Eduardo Martins]

Julian, there is not nothing private in Mobicent, this the truly open
and public community :)

Now such extensive load test report was not repeated lately, due to
higher priorities, but if you want to know about JAIN SLEE
specifically, latest release does around 1100 cps in same test,
including same hardware.

-- Eduardo
..............................................
http://emmartins.blogspot.com
http://redhat.com/solutions/telco

[Julian]

Hope you have some time to release some tests like before.

1100 cps is really nice btw! Congrats.

Best regards.

2011/12/2 Eduardo Martins <emma...@gmail.com>

[aayush]

When you try the next perf test, try including this flag too in JAVA_OPTS-

-XX:+UseNUMA

This flag has been available since JDK 6 Update 2, and it works with Parallel GC flag enabled.

By using the NUMA flag, JDK performance doc says that performance increases by approx 30% on 32-bit systems and by approx 40% on 64-bit systems w.r.t. Spec4j benchmark.

I dont know how true this is, but its worth a try.

Regards

Aayush

On Dec 3, 2011 7:38 AM, "Julian" <caci...@gmail.com> wrote:

Hope you have some time to release some tests like before.

1100 cps is really nice btw! Congrats.

Best regards.

2011/12/2 Eduardo Martins <emma...@gmail.com>
>
> Julian, there is not nothing private in Mobice...