Hey guys,
Just wondering if anyone could explain a difference in benchmark numbers between iometer and SQLIO.
To set this question up:
I'm trying to get baseline performance for a new SAN installation, and have been testing with SQLIO for a particular workload: Random 8kB to 32kB file access (10:1 read:write) against a fileshare containing millions of these small files. The numbers were looking pretty good, example: 8k-read-random, 8-threads, 8 outstanding requests, hwbuffer, 120sec resulted in 100k IOPS.
A colleague suggested using iometer to confirm results, but when I configure iometer for same io profile and test params, I get lower figures. So for the example above, I get about 60k IOPS in iometer.
The physical disk performance counters match up with during SQLIO and iometer tests, so I am confident that the results produced by each are valid. I'm looking at average request size (to confirm size of io request), disk transfers per second, amongst others.
I'm just not sure why there is such a disparity... To be specific, I can see 100k disk requests/sec in perfmon, 100k IOPS in the SQLIO results output, and 100k IOPS in the storage vendor's separate reporting tool. I've double checked configuration parameters between the two, and played with iometer values (like queue size or worker count) to see if I could coax greater numbers out of it. Haven't been able to yet.
Who to believe? Is iometer less capable of pushing IO, or is it the more accurate one? Hard to see SQLIO not being "accurate" since Microsoft's teams use it themselves.
Thanks!