Completing and merging core scheduling

You are correct that linpack alone doesn't have this issue when the system is not overcommitted (and actually benefit from smt off). However, in this case, it is running concurrently with 96 1-vcpu noise VMs on the same NUMA node (3:1 overcommit that becomes 6:1 with smt off). I clarified this above for the MySQL benchmark, but the linpack benchmark was running in the same configuration.