Bug #9076

ZFS Performance test concurrency should be lowered for better latency

Added by John Kennedy 12 days ago.

Status:NewStart date:2018-02-07
Priority:NormalDue date:
Assignee:John Kennedy% Done:


Target version:-
Difficulty:Medium Tags:needs-triage


Work by Stephen Blinick.

Nightly performance runs typically consist of two levels of concurrency; and both are fairly high.

Since the IO runs are to a ZFS filesystem, within a zpool, which is based on some variable number of vdev's, the amount of IO driven to each device is variable. Additionally, different device types (HDD vs SSD, etc) can generally handle a different amount of concurrent IO before saturating.

Nevertheless, in practice, it appears that most tests are well past the concurrency saturation point and therefore both perform with the same throughput, the maximum of the device. Because the queuedepth to the device(s) is so high however, the latency is much higher than the best possible at that throughput, and increases linearly with the increase in concurrency.

This means that changes in code that impact latency during normal operation (before saturation) may not be apparent when a large component of the measured latency is from the IO sitting in a queue to be serviced. Therefore, changing the concurrency settings is recommended

Also available in: Atom