Nick Piggin [interview] updated his -np patchset against 2.6.8.1-mm2, introducing "per-zone and dchache scanning lists", and "'free_local_harder', which attempts to free some node local memory before going remote." Nick explains, "these should be good for NUMA performance, although free_local_harder will, by definition, introduce regressions for some workloads."
From: Nick Piggin [email blocked] To: linux-kernel [email blocked] Subject: 2.6.8.1-np1 Date: Fri, 20 Aug 2004 18:27:45 +1000 http://www.kerneltrap.org/~npiggin/2.6.8.1-np1/ Again, this is only against 2.6.8.1-mm2 for now due to the amount of scheduler and memory management patches in Andrew's tree. This introduces per-zone inode and dcache scanning lists. These introduce a theoretical problem (solveable, but would take some work) however in practice I don't think it would cause a problem. Also introduces "free_local_harder", which attempts to free some node local memory before going remote. It is quite dumb at the moment, and performs the scanning from process context (should be moved to kswapd context), however it seems to usually do what it is supposed to. It can be switched off by echo 0 > /proc/sys/vm/free_local_harder. These should be good for NUMA performance, although free_local_harder will, by definition, introduce regressions for some workloads. Unfortunately, free_local_harder sits pretty firmly on top of my other memory management changes, which may introduce problems themselves. Even so, feedback would be cool.
