Benchmarking running Swingbench
Posted by John Hallas on March 4, 2008
Swingbench is an excellent toolset and we are pushing our test rig by using it. However I am not sure how to tell when you have an optimum configuration or what a realistic throughput (TPM might be).
The best source of information I have seen is a paper by Larry Pedigo of Oracle on a Linux/Windows comparison using Swingbench.
In that, using a 2 node RH4 pair of HP DL380 servers and a SAN back end he sees a transaction per minute (TPM) of around 76000 (according to a screen shot which I think was with his optimum configuration).
We are using Sun V585 servers with OEL5 and a solid state disk backend. Our figures show a remarkable similarity with a constant level of around 75000 TPM. Oracle 11g
Both tests referred to above were using the Order Entry stress test with 2 servers generating load and the configuration set to 125 users per generator and 0 think and wait time (pretty much as Larry covers in his paper).
When we ran an insert stress test (again using the Order Entry schema) we achieved 410000 TPM with a single load generator to a single instance, 320000 when both instances were running and a pathetic 120000 when we added a second load generator.
Under every test we see the main problem being the major contention on gc_buffer_busy blocks.
We did get benefits on binding a second interface card and adding that to the CRS config but when we added a third card that seemed to reduce throughput and we had several node crashes.
I am not really asking a question here but hopefully these notes may be useful for others when they are searching for Swingbench throughput rates and results. Any comments appreciate though