Be aware that on larger cloud vendor instances, the backing vCPUs and memory you get allocated might or might not be spread opaquely across multiple sockets and/or NUMA nodes of the backing hardware. Some of them have options where you can allocate raw hardware as well so you can opt to lock your process to run within just one NUMA node and ensure hardware locality for your performance debugging.

I'm pointing this out in case you experience any mystery jitters in your benchmark results.