On Tue, Mar 07, 2017 at 02:15:37PM +0100, Vlastimil Babka wrote: > Last year, Johannes Weiner has reported a regression in page mobility > grouping [1] and while the exact cause was not found, I've come up with some > ways to improve it by reducing the number of allocations falling back to > different migratetype and causing permanent fragmentation. I finally managed to get a handful of our machines on 4.10 with these patches applied and a 4.10 vanilla control group. The sampling period is over twelve hours, which is on the short side for evaluating that load, so take the results with a grain of salt. The allocstall rate (events per second) is down on average, but there are occasionally fairly high spikes that exceed the peaks in 4.10: http://cmpxchg.org/antifrag/allocstallrate.png Activity from the compaction free scanner is down, while the migration scanner does more work. I would assume most of this is coming from the same-migratetype restriction on the source blocks: http://cmpxchg.org/antifrag/compactfreescannedrate.png http://cmpxchg.org/antifrag/compactmigratescannedrate.png Unfortunately, the average compaction stall rate is consistently much higher with the patches. The 1h rate averages are 2-3x higher: http://cmpxchg.org/antifrag/compactstallrate.png An increase in direct compaction is a bit worrisome, but the task completion rates - the bottom line metric for this workload - are still too chaotic to say whether the increased allocation latency affects us meaningfully here. I'll give it a few more days. Is there any other data you would like me to gather? Thanks! -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
