Benchmark of BEA WebLogic SIP Server on Intel Xeon Processor
Pages: 1, 2
These tests were performed in July 2007 at Intel's Embedded and Communications Solutions lab in Kontich, Belgium, by Intel engineers with the support of BEA. The following table summarizes the results achieved for the different call durations tested.
|Call Duration||No. of JVM||Processor||CPS||CPU Utilization|
As one would expect, the load that the system can sustain while remaining within the allowed operational limits decreases as the call duration increases. This is due to the higher number of active sessions, which causes an increase in the heap usage and requires more work from the SIP server when dispatching messages to their session contexts. The decrease, however, is not dramatic considering the fairly large memory footprint of SIP session objects (30-50 KB).
As can be seen from the CPU utilization numbers, there should be enough headroom to add some flesh to the application without significantly impacting the achieved call rate.
It should also be noted that the performance achieved by this Java-based implementation is not so far from what is typically seen in native (C/C++) SIP proxy implementations on similar platforms.
The graphs included in this section show, for the various call durations, the evolution of the call setup time and the CPU utilization as the call rate increases. The retransmission and failure rates are also shown.
With calls that are immediately disconnected, the CPU usage increases linearly with the call rate.
With SIP sessions that remain active for 40 seconds, dispatching messages to the appropriate session handler requires more CPU cycles, as well as managing a larger Java heap. This is probably what explains the fact that the CPU utilization is now increasing more than linearly with the call rate.
At 80s call duration, the trend that was seen with 40s call duration is just reinforced.
This initial round of benchmark performance figures demonstrates that a BEA WebLogic Communication Platform solution is capable of handling in excess of 1000 SIP calls per second (CPS) while maintaining a sufficiently low CPU utilization (approximately 50 percent), and still providing headroom for deployment of additional Java applications or introducing high availability features such as clustering and replication.
By scaling up the number of BEA WebLogic SIP Server instances and utilizing the clustering and replication capabilities of this platform, in combination with the performance and reliability of Intel Xeon Processor 5100 and 5300 Series, there is still enormous scope for further improvement. This demonstrates that Intel and BEA are a fantastic combination and should form the foundation for application and solution deployment.
This also clearly demonstrates that Java-based solutions are able to handle the high throughput and low latency requirements required of today's telco networks.
The authors would like to thank Francois Deza, Senior Manager Business Development BEA Systems, and Stefan Sarne, Senior Engineer BEA Systems, for their contribution to the article and for the support during the test phase.
David Verbeiren is an Application Engineer within Intel's Embedded and Communications Group.