As you all know GigaSpaces XAP 6.5 is on its way out to the market. This release involves incredible effort that is basically a collection of large amount of improvements with the product scalability. These are result of feedback we gathered from the field and customers around the globe. One of the final tasks we have done was measuring the product performance in different aspects where the main focus is to measure how far it can scale.
The focus was on 2 main tests:
– Scale up (vertically aka in-process) – This means the ability to serve more concurrent application threads vs., increase with the overall throughput (theoretically in linear manner) until all CPU and memory resources fully consumed.
– Scale out (horizontally aka out-of-process) – This means the ability to serve more concurrent application processes vs., increase with the overall throughput (theoretically in linear manner) until all CPU and memory resources fully consumed. In other words: adding additional hardware resources increase the overall capacity in linear manner.
We used mainly Sun HW, running Linux/Sun OS with Sun regular and Real Time JVM . For extreme in-proc scalability tests, we have been using Azul Vega 3 appliance with its 207 cores. The work conducted with cooperation with Sun , Novell , Mellanox , Voltaire , Israeli Association of Grid Technologies, and Azul. Thank you guys for your great help!
The product demonstrated great scalability, stability and robustness , with all the various scenarios it has been tested with. In order to simulate real life usage , we have been using 99% of the out of the box default settings with very minor config optimization. No special product config , no special JVM optimizations , or special OS optimization been done. We have been doing also some Infiniband tests shows also interesting improvements with space operations with large objects.
Another interesting results can be seen with the deterministic behavior of the Sun real time JVM ).
The Java real time benchmark running an embedded space with 10 threads performing parallel write and take operations in a rate of 200 oper/sec i.e. 2000 write and 2000 take per second.
The test run 25 minutes performing 3,000,000 write operations and 3,000,000 take operations.
The results show very steady behavior. The throughput is stable without oscillation around the target throughput.
If you are interested with a copy of the benchmark report just drop us a line at: sales at giagspaces dot com.