sched/numa: Update migrate_improves/degrades_locality()
authorRik van Riel <riel@redhat.com>
Thu, 15 May 2014 17:03:06 +0000 (13:03 -0400)
committerIngo Molnar <mingo@kernel.org>
Thu, 22 May 2014 09:16:39 +0000 (11:16 +0200)
commitb1ad065e65f56103db8b97edbd218a271ff5b1bb
treea44ca4ffbfc2fcc618ac668daffda4ef57f4988e
parente63da03639cc9e6e83b62e7ef8ffdbb92421416a
sched/numa: Update migrate_improves/degrades_locality()

Update the migrate_improves/degrades_locality() functions with
knowledge of pseudo-interleaving.

Do not consider moving tasks around within the set of group's active
nodes as improving or degrading locality. Instead, leave the load
balancer free to balance the load between a numa_group's active nodes.

Also, switch from the group/task_weight functions to the group/task_fault
functions. The "weight" functions involve a division, but both calls use
the same divisor, so there's no point in doing that from these functions.

On a 4 node (x10 core) system, performance of SPECjbb2005 seems
unaffected, though the number of migrations with 2 8-warehouse wide
instances seems to have almost halved, due to the scheduler running
each instance on a single node.

Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: mgorman@suse.de
Cc: chegu_vinod@hp.com
Link: http://lkml.kernel.org/r/20140515130306.61aae7db@cuia.bos.redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/fair.c