Completing and merging core scheduling
Completing and merging core scheduling
Posted May 13, 2020 18:03 UTC (Wed) by jdesfossez (subscriber, #123831)Parent article: Completing and merging core scheduling
> MySQL benchmark showing performance dropping by 60% when core scheduling is used
Just a quick precision, this test case was performed with 96 1-vcpu "noise VMs" running at the same time on the same NUMA (a 3:1 overcommit ratio).
So yes it's a significant drop, but it is on a very busy host running some kind of a worst case scenario with lots of 1-vcpu VMs. Core scheduling (or nosmt for that matter) doesn't show a significant performance degradation if there is enough headroom left on the system.
Just a quick precision, this test case was performed with 96 1-vcpu "noise VMs" running at the same time on the same NUMA (a 3:1 overcommit ratio).
So yes it's a significant drop, but it is on a very busy host running some kind of a worst case scenario with lots of 1-vcpu VMs. Core scheduling (or nosmt for that matter) doesn't show a significant performance degradation if there is enough headroom left on the system.