Some takeaways from the Sun/Oracle TPC-C benchmark
While reading the Full Disclosure reports of Suns TPC-C benchmark report i found some interesting things:
- The average response times in the Sun TPC-C result are 17 times better on average than the one in IBM's result.
- OpenSolaris 2009.06 was central to the benchmark. As a Sun Storage F5100 isn't capable of multipathing and multinitator capable in itself (for a single zone), the COMSTAR SCSI target framework was used to provide access to the F5100 storage arrays. In this case the targets were presented via 8 GBit/s FC. Nice to see OpensSolaris in such an important role in a configuration.
- As far as i'm understanding the documentation of the Sun Benchmark, the tablespaces were configured with a blocksize of 4k. At this block size the system yields 1.6M read IOPS and 1.2M write IOPS. But as specified by the the list of material each F5100 was just connected to 4 HBA. This is the so-called "HBA economy" configuration. You have to know that a single SAS HBA connection delivers "just" 150.000 IOPS. So each F5100 could be loaded just to 600.000 IOPS. Would be nice to see the TPC-C numbers while using 4 systems with 4 HBA to use the F5100 in the maximum performance configuration. Looks like the F5100 fought with one arm tied to the back. I think we can expect some additional news from this bunch of F5100 devices ;)
- Interestingly the benchmark documentation counts every FMod as a disk. Thus this benchmark used 5657 disks. But that is a number where IBMs result still leads. They used 11.000 for their result.