Details
-
Improvement
-
Resolution: Fixed
-
Major
-
None
-
None
-
3
-
9223372036854775807
Description
Recently running fio tests with certain processors has been showing more than 50% performance degradation compared with previous processor generation.
Doing more investigations has proven that the problem was caused by the latency to exit from idle state when some deeper sleeping/saving power-management modes were configured.
In recent (6.8+) kernels, there are already several drivers preventing this behaviour by using the PM-QoS Linux kernel framework and ensuring to "bump" the power-management mode to guarantee a minimal latency during on-going IOs/operations, for the CPUs/Cores that will be involved to handle completions.
We have done some similar instrumentations in Lustre-client code and have been able to recover the expected performance.
I am presently porting these changes to master and will push a patch asap.