Tests methodology

Each single disk test returns at least three numbers - bandwith, IOPS and latency. There a lot of parameters, which affect this numbers. Apart from system setup itself next parameters are part of test load:

• Block size. Power of two [512, 4 MiB]
• Operation - {initial write, read, write, read+write}
• Offset selection mode - {sequential, random}
• Sync mode - {sync, direct, async, direct+sync}
• VM count - [1, several dozens] typically
• Client count - [1, up to thousand] per VM
• Test len - X sec, [10s, Several hours]

Run long enough tests for each possible combinations would requires a weeks of testing. We would try to cut those tests, which are:

• Depends on other parts of system more, that on disk storage
• Don't give new data, as they results can be estimated from results of other tests

For different type of testing system there would be a different tests selection:

• HDD or local HDD RAID
• SSD or local SSD RAID
• Ceph
• Cinder over ISCSI