Messages in this thread | | | Date | Wed, 28 Aug 2013 11:32:19 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH 1/8] THP: Use real address for NUMA policy |
| |
On Wed, Aug 28, 2013 at 11:18:14AM +0200, Ingo Molnar wrote: > But ideally THP should detect cases where a hugepage is heavily used from > multiple nodes and un-HP the page in question. Or not turn it into a > hugepage in the first place. (We actually have a memory access pattern > sampling facility in place upstream, the new CONFIG_NUMA_BALANCING=y code, > which could in theory be utilized.)
Mel and I spoke about doing just that a few weeks ago. Breaking up THP pages when we get shared faults from different nodes and avoiding coalescing when the pages are from different nodes.
Rik seemed a little skeptical about the entire thing saying that the huge tlb gain is about the same order as the locality thing -- while this is more or less true for the 'tiny' machines most of us have, this is clearly not the case for these big machines.
| |