Serial ATA performance IOMeter
We'll begin our storage tests with IOMeter, which subjects our systems to increasing multi-user loads. We used IOMeter's workstation and database test patterns, since those are more relevant to desktop systems than the file or web server test patterns. This particular test makes good use of the Native Command Queuing capability built into the AHCI specification.
Drives capable of taking advantage of the SB850's 6Gbps SATA controller are few and far between. The best candidate is currently Crucial's new RealSSD C300, which is the first 6Gbps solid-state drive to hit The Benchmarking Sweatshop. Naturally, we couldn't resist testing with it. But plenty of folks also use mechanical drives, and likely will for some time, so we'll kick things off with a look at SATA performance using a Western Digital VelociRaptor.
AMD has a history of poor storage controller drivers, so in addition to testing the Gigabyte 890GX board with AMD's own drivers, we tested it with the Microsoft AHCI drivers included with Windows 7.
When they're both using AMD's AHCI drivers, the Asus and Gigabyte 890GX boards offer nearly identical transaction rates. Performance levels off after we hit 32 concurrent I/O requests, which just happens to be the queue depth for Native Command Queuing. Interestingly, there's no performance plateau when we combine the 890GX with Microsoft's own AHCI drivers. The H55 Express' transaction rates don't trail off after 32 I/Os, either.
Regardless of which drivers are used, the Gigabyte 890GX board uses relatively more CPU time than the Asus. The H55's CPU utilization is lower than both boards, although we're still looking at less than 4% CPU utilization overall.
Switching over to a high-end SSD exposes a weakness with Microsoft's AHCI drivers, at least at lower loads. The H55 Express still manages slightly higher transaction rates than any of our 890GX configurationsand that's with a 3Gbps SATA controller.
Again, the Gigabyte 890GX board exhibits higher CPU utilization than the Asus. Both consume relatively more CPU cycles than our H55 system, but the differences don't amount to much as the load scales upward. I didn't expect CPU utilization to peak with the fewest outstanding I/O requests, but that's what happened with all four configurations.