To use Ceph's built-in tool, first create a dedicated, disposable pool to scribble into. For this to be a legitimate test, it must have the same attributes as your production pool: PG count, replication factor, CRUSH rules, etc.

Options to rados bench include:

• -p: The name of our dedicated test pool
• seconds: Number of seconds the test should run
• write|seq|rand: Type of workload to present: write, sequential read, random read
• -t: Number of concurrent operations; the default is 16
• --no-cleanup: Don't clean up the objects created during the run

Let's run a 60-second write test against a small cluster with a pool named data. A dedicated pool should be used to ensure that user data is not clobbered. Longer tests are better than shorter, as they ensure that caching effects do not overly skew the results:

# rados bench -p data 60 write
Maintaining 16 concurrent writes of 4194304 bytes for up to 60 seconds or 0 objects
Object prefix: benchmark_data_ncc_1701
sec Cur ops started finished avg MB/s cur MB/s last lat avg lat
0 0 0 0 0 0 - 0
1 16 39 23 92.9777 93 0.627 0.56
2 16 67 50 98.853 109 0.523 0.59
3 16 97 80 105.234 119 0.267 0.55
4 16 126 109 107.843 114 0.231 0.54
5 16 152 135 106.111 103 0.301 0.52
6 16 181 164 108.334 114 0.714 0.56
7 16 209 192 108.945 110 0.843 0.59
8 16 237 220 108.238 111 0.133 0.53
9 16 266 249 109.934 113 0.780 0.53
10 15 292 276 109.364 111 0.822 0.51
...

Tests should be run multiple times and for longer periods and the results compared and averaged to validate the method.