3 ms·
Scheduling threads is probably a big one I imagine. If Windows schedules two threads from the same process working on a shared dataset across two different NUMA
by magicalhippo 8y ago
Scheduling threads is probably a big one I imagine. If Windows schedules two threads from the same process working on a shared dataset across two different NUMA nodes then memory latency can shoot up, from what I understand.
I've also noticed that Windows is very fond of passing hard-working threads around all the cores in turn, rather than letting them hog certain cores. Pretty sure that doesn't improve cache behavior etc.