On 3/23/22 01:41, Paul Gofman wrote:
On 3/22/22 19:37, Jinoh Kang wrote:
In my experience on Linux pre-faulting the memory this way reduce the overhead greatly, while the benefit of explicit huge pages management over this is not apparent. That shows stark difference from my experience. In my case, memset()-ing yielded no significant improvement in performance (the faults moves from blitting to zeroing).
Did you have transparent huge pages enabled?
Sorry, I somehow misread "explicit HP" as "transparent HP." It now makes total sense. To answer your question, no (for some reason). I think it's safe to treat myself as a special case here.
Not all platforms have THP though, so I'll have to evaluate whether my patchset could bring more advantage at less VM overhead on such platforms.