Guessing this is especially hard to automate with peripherals involved. I once had a workload slow severely because it was running on the NUMA node that didn't share memory with the NIC.