Messages in this thread | | | Date | Tue, 22 Sep 2009 11:05:40 +0100 | From | Mel Gorman <> | Subject | Re: [RFC PATCH 0/3] Fix SLQB on memoryless configurations V2 |
| |
On Mon, Sep 21, 2009 at 02:17:40PM -0400, Christoph Lameter wrote: > On Mon, 21 Sep 2009, Mel Gorman wrote: > > Can you spot if there is something fundamentally wrong with patch 2? I.e. what > > is wrong with treating the closest node as local instead of only the > > closest node? > > Depends on the way locking is done for percpu queues (likely lockless). > A misidentification of the numa locality of an object may result in locks > not being taken that should have been taken. >
Ok, I'll continue looking from that perspective and see what comes out. I've spotted a few possible anomolies which I'll stick into a separate patch.
> > > Or just allow SLQB for !NUMA configurations and merge it now. > > > > > > > Forcing SLQB !NUMA will not rattle out any existing list issues > > unfortunately :(. > > But it will make SLQB work right in permitted configurations. The NUMA > issues can then be fixed later upstream. >
I'm going to punt the decision on this one to Pekka or Nick. My feeling is leave it enabled for NUMA so it can be identified if it gets fixed for some other reason - e.g. the stalls are due to a per-cpu problem as stated by Sachin and SLQB happens to exasperate the problem.
-- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
| |