Ceph tests selection

Report example See How-to-interpret-performance-reports for interpretation.

Test starts(as all other tests) from filling test file with pseudo-random data, generated by fio. Than all test, described below executed. Wally analyze fio IOPS/BW/Lat logs to calculate deviations and confidence intervals.

Next tests are included into default ceph suite.

Common parameters:

• ramp_time=30
• runtime=180

Tests:

• randwrite, 4KiB, sync, th_count={1, 5, 10, 15, 25, 40}
• randwrite, 4KiB, direct, th_count=1
• randread, 4KiB, direct, th_count={1, 5, 10, 15, 25, 40, 80, 120}
• randwrite, 16MiB, direct, th_count={1, 3, 10}
• randread, 16MiB, direct, th_count={1, 3, 10}

Selection reasoning

Tests measure:

• 4k Read and write IOPS with different amount of threads. Write in sync mode
• 4k direct write, as for direct write even single thread creates maximum load
• Linear read/write. Linear read/writes in case of ceph can scale with threads. Ceph split data for rbd device on 4M block, so 16M random is equals to "linear write". "Real" linear writes from different threads into the same block would be merged and results performance would be wrong. Also due to fio limitation it's impossible to run different threads from different offset easily.