On Wed, Sep 29, 2010 at 09:34:09AM -0500, Christoph Lameter wrote: > On Wed, 29 Sep 2010, Mel Gorman wrote: > > > > Updating the threshold also is expensive. > > > > Even if it's moved to a read-mostly part of the zone such as after > > lowmem_reserve? > > The threshold is stored in the hot part of the per cpu page structure. > And the consequences of moving it? In terms of moving, it would probably work out better to move percpu_drift_mark after the lowmem_reserve and put the threshold after it so they're at least similarly hot across CPUs. > > > I thought more along the lines > > > of reducing the threshold for good if the VM runs into reclaim trouble > > > because of too high fuzziness in the counters. > > > > > > > That would be unfortunate as it would only take trouble to happen once > > for performance to be impaired for the remaining uptime of the machine. > > Reclaim also impairs performance and inaccurate counters may cause > unnecessary reclaim. Ah, it's limited to be fair. You might end up reclaiming "maximum drift" number of pages you didn't need to but that doesn't seem as bad. > Ultimately this is a tradeoff. The current thresholds > were calculated so that there will be zero impact even for very large > configurations where all processors continual page fault. I think we have > some leeway to go lower there. The tuning situation was a bit extreme. > Ok. -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>