Messages in this thread | | | Date | Wed, 12 Sep 2018 16:15:05 +0530 | From | Srikar Dronamraju <> | Subject | Re: [PATCH 4/4] sched/numa: Do not move imbalanced load purely on the basis of an idle CPU |
| |
* Peter Zijlstra <peterz@infradead.org> [2018-09-12 11:36:21]:
> On Wed, Sep 12, 2018 at 12:24:10PM +0530, Srikar Dronamraju wrote: > > > Kernel A = 4.18+ 13 sched patches part of v4.19-rc1. > > Kernel B = Kernel A + 6 patches (http://lore.kernel.org/lkml/1533276841-16341-1-git-send-email-srikar@linux.vnet.ibm.com) > > Kernel C = Kernel B - (Avoid task migration for small numa improvement) i.e > > http://lore.kernel.org/lkml/1533276841-16341-4-git-send-email-srikar@linux.vnet.ibm.com > > + 2 patches from Mel > > (Do not move imbalanced load purely) > > http://lore.kernel.org/lkml/20180907101139.20760-5-mgorman@techsingularity.net > > (Stop comparing tasks for NUMA placement) > > http://lore.kernel.org/lkml/20180907101139.20760-4-mgorman@techsingularity.net > > But that's not a fair comparison :/ You've complicated things by adding > that second patch from Mel. >
One patch does choose a idle thread over a possible swap. This is the one that conflicts with "Avoid task migration for small numa improvement".
The other one detects if we have already found an idle thread and stops iterating. So if we have already chosen a idle CPU, then dont iterate. So this patch should ideally help the above patch.
My take on both the patches: If there is a possible swap candidate (i.e if choosing a thread over idle is better from a NUMA score improvement perspective), then it means the swap candidate will benefit from movement. It means the swap candidate is not running on its preferred node at this moment. If we don't do it, the swap candidate will later on try migrating to its preferred node. We could avoid that if we choose swap candidate over idle thread. Please note if the idle and swap candidates yield the same NUMA score, then we still choose the idle thread and not the swap candidate.
> Now you cannot (unambiguously) tell what the cause for the performance > difference is. >
The difference as I listed above is should we choose a swap candidate with more NUMA score compared to idle thread.
My patch would hurt if there was an idle thread and we still iterate for possible swap candidates and we dont find any. But that we cant speculate.
| |