Messages in this thread | | | Subject | Re: [PATCH 13/40] autonuma: CPU follow memory algorithm | From | Peter Zijlstra <> | Date | Fri, 29 Jun 2012 20:51:45 +0200 |
| |
On Fri, 2012-06-29 at 14:46 -0400, Rik van Riel wrote: > > I've also stated several times that forceful migration in the context of > > numa balancing must go. > > I am not convinced about this part either way. > > I do not see how a migration numa thread (which could potentially > use idle cpu time) will be any worse than migrate on fault, which > will always take away time from the userspace process.
Any NUMA stuff is long term, it really shouldn't matter on the timescale of a few jiffies.
NUMA placement should also not over-ride fairness, esp. not by default.
| |