Messages in this thread | | | Date | Thu, 02 Oct 2014 15:28:39 -0400 | From | Rik van Riel <> | Subject | Re: [PATCH 4/4] mm: numa: Do not mark PTEs pte_numa when splitting huge pages |
| |
On 10/02/2014 03:26 PM, Linus Torvalds wrote: > On Thu, Oct 2, 2014 at 12:07 PM, Kirill A. Shutemov > <kirill@shutemov.name> wrote: >> >> I believe Sasha uses fakenuma in his KVM for that. > > Ok, so the benchmarks won't do anything then. > > I mean, I guess they might show some of the migration overhead, but > they won't show the actual end result in any meaningful manner, since > memory isn't actually NUMA.
Both autonuma and "perf bench numa mem" mostly tell us how quickly the kernel manages to locate tasks and their memory on the nodes where they belong, without doing much in the way of NUMA performance measuring.
They are more useful as sanity tests than anything else.
"Does the kernel still properly place each process on its own node, and how quickly does it do that?"
| |