On 03/16/2017 07:34 PM, Johannes Weiner wrote: > On Wed, Mar 08, 2017 at 08:17:39PM +0100, Vlastimil Babka wrote: >> On 8.3.2017 17:46, Johannes Weiner wrote: >>> Is there any other data you would like me to gather? >> >> If you can enable the extfrag tracepoint, it would be nice to have graphs of >> how >> unmovable allocations falling back to movable pageblocks, etc. > > Okay, here we go. I recorded 24 hours worth of the extfrag tracepoint, > filtered to fallbacks from unmovable requests to movable blocks. I've > uploaded the plot here: > > http://cmpxchg.org/antifrag/fallbackrate.png > > but this already speaks for itself: > > 11G alloc-mtfallback.trace > 3.3G alloc-mtfallback-patched.trace > > ;)
Great! >> Possibly also /proc/pagetypeinfo for numbers of pageblock types. > After a week of uptime, the patched (b) kernel has more movable blocks > than vanilla 4.10-rc8 (a): > > Number of blocks type Unmovable Movable Reclaimable HighAtomic > CMA Isolate > > a: Node 1, zone Normal 2017 29763 987 1 > 0 0 > b: Node 1, zone Normal 1264 30850 653 1 > 0 0 That's better than I expected. I wouldn't be surprised if the number of unmovable pageblocks actually got *higher* due to the series because previously many unmovable pages would be scattered around movable blocks. > I sampled this somewhat sporadically over the week and it's been > reading reliably this way. > > The patched kernel also consistently beats vanilla in terms of peak > job throughput. > > Overall very cool! Thanks a lot! So that means it's worth the increased compaction stats you reported earlier?