| Message ID | 20181214230310.572-10-mgorman@techsingularity.net (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Increase success rates and reduce latency of compaction v1 | expand |
On 12/15/18 12:03 AM, Mel Gorman wrote: > When pageblocks get fragmented, watermarks are artifically boosted to pages > are reclaimed to avoid further fragmentation events. However, compaction > is often either fragmentation-neutral or moving movable pages away from > unmovable/reclaimable pages. As the actual watermarks are preserved, > allow compaction to ignore the boost factor. Right, I should have realized that when reviewing the boost patch. I think it would be useful to do the same change in __compaction_suitable() as well. Compaction has its own "gap". > 1-socket thpscale > 4.20.0-rc6 4.20.0-rc6 > finishscan-v1r4 noboost-v1r4 > Amean fault-both-1 0.00 ( 0.00%) 0.00 * 0.00%* > Amean fault-both-3 3849.90 ( 0.00%) 3753.53 ( 2.50%) > Amean fault-both-5 5054.13 ( 0.00%) 5396.32 ( -6.77%) > Amean fault-both-7 7061.77 ( 0.00%) 7393.46 ( -4.70%) > Amean fault-both-12 11560.59 ( 0.00%) 12155.50 ( -5.15%) > Amean fault-both-18 16120.15 ( 0.00%) 16445.96 ( -2.02%) > Amean fault-both-24 19804.31 ( 0.00%) 20465.03 ( -3.34%) > Amean fault-both-30 25018.73 ( 0.00%) 20813.54 * 16.81%* > Amean fault-both-32 24380.19 ( 0.00%) 22384.02 ( 8.19%) > > The impact on the scan rates is a mixed bag because this patch is very > sensitive to timing and whether the boost was active or not. However, > detailed tracing indicated that failure of migration due to a premature > ENOMEM triggered by watermark checks were eliminated. > > Signed-off-by: Mel Gorman <mgorman@techsingularity.net> > --- > mm/page_alloc.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 80535cd55a92..c7b80e62bfd9 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -3043,7 +3043,7 @@ int __isolate_free_page(struct page *page, unsigned int order) > * watermark, because we already know our high-order page > * exists. > */ > - watermark = min_wmark_pages(zone) + (1UL << order); > + watermark = zone->_watermark[WMARK_MIN] + (1UL << order); > if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA)) > return 0; > >
On Tue, Dec 18, 2018 at 01:36:42PM +0100, Vlastimil Babka wrote: > On 12/15/18 12:03 AM, Mel Gorman wrote: > > When pageblocks get fragmented, watermarks are artifically boosted to pages > > are reclaimed to avoid further fragmentation events. However, compaction > > is often either fragmentation-neutral or moving movable pages away from > > unmovable/reclaimable pages. As the actual watermarks are preserved, > > allow compaction to ignore the boost factor. > > Right, I should have realized that when reviewing the boost patch. I > think it would be useful to do the same change in > __compaction_suitable() as well. Compaction has its own "gap". > That gap is somewhat static though so I'm a bit more wary of it. However, the check in __isolate_free_page looks too agressive. We isolate in units of COMPACT_CLUSTER_MAX yet the watermark check there is based on the allocation request. That means for THP that we check if 512 pages can be allocated when only somewhere between 1 and 32 is needed for that compaction cycle to complete. Adjusting that might be more appropriate?
On 12/18/18 2:51 PM, Mel Gorman wrote: > On Tue, Dec 18, 2018 at 01:36:42PM +0100, Vlastimil Babka wrote: >> On 12/15/18 12:03 AM, Mel Gorman wrote: >>> When pageblocks get fragmented, watermarks are artifically boosted to pages >>> are reclaimed to avoid further fragmentation events. However, compaction >>> is often either fragmentation-neutral or moving movable pages away from >>> unmovable/reclaimable pages. As the actual watermarks are preserved, >>> allow compaction to ignore the boost factor. >> >> Right, I should have realized that when reviewing the boost patch. I >> think it would be useful to do the same change in >> __compaction_suitable() as well. Compaction has its own "gap". >> > > That gap is somewhat static though so I'm a bit more wary of it. However, Well, watermark boost is dynamic, but based on allocations stealing from other migratetypes, not reflecting compaction chances of success. > the check in __isolate_free_page looks too agressive. We isolate in > units of COMPACT_CLUSTER_MAX yet the watermark check there is based on > the allocation request. That means for THP that we check if 512 pages > can be allocated when only somewhere between 1 and 32 is needed for that > compaction cycle to complete. Adjusting that might be more appropriate? AFAIU the code in __isolate_free_page() reflects that if there's less than 512 free pages gap, we might form a high-order page for THP but won't be able to allocate it afterwards due to watermark.
On Tue, Dec 18, 2018 at 02:58:33PM +0100, Vlastimil Babka wrote: > On 12/18/18 2:51 PM, Mel Gorman wrote: > > On Tue, Dec 18, 2018 at 01:36:42PM +0100, Vlastimil Babka wrote: > >> On 12/15/18 12:03 AM, Mel Gorman wrote: > >>> When pageblocks get fragmented, watermarks are artifically boosted to pages > >>> are reclaimed to avoid further fragmentation events. However, compaction > >>> is often either fragmentation-neutral or moving movable pages away from > >>> unmovable/reclaimable pages. As the actual watermarks are preserved, > >>> allow compaction to ignore the boost factor. > >> > >> Right, I should have realized that when reviewing the boost patch. I > >> think it would be useful to do the same change in > >> __compaction_suitable() as well. Compaction has its own "gap". > >> > > > > That gap is somewhat static though so I'm a bit more wary of it. However, > > Well, watermark boost is dynamic, but based on allocations stealing from > other migratetypes, not reflecting compaction chances of success. > True. > > the check in __isolate_free_page looks too agressive. We isolate in > > units of COMPACT_CLUSTER_MAX yet the watermark check there is based on > > the allocation request. That means for THP that we check if 512 pages > > can be allocated when only somewhere between 1 and 32 is needed for that > > compaction cycle to complete. Adjusting that might be more appropriate? > > AFAIU the code in __isolate_free_page() reflects that if there's less > than 512 free pages gap, we might form a high-order page for THP but > won't be able to allocate it afterwards due to watermark. Yeah but it used to be a lot more important when watermark checking for high-orders was very different. Now, if the watermark is met for order-0 and a large enough free page is allocated, the allocation succeeds so it's a lot less relevant than it used to be. kswapd will still run in the background for order-0 if necessary so a heavy watermark check there doesn't really help.
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 80535cd55a92..c7b80e62bfd9 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -3043,7 +3043,7 @@ int __isolate_free_page(struct page *page, unsigned int order) * watermark, because we already know our high-order page * exists. */ - watermark = min_wmark_pages(zone) + (1UL << order); + watermark = zone->_watermark[WMARK_MIN] + (1UL << order); if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA)) return 0;
When pageblocks get fragmented, watermarks are artifically boosted to pages are reclaimed to avoid further fragmentation events. However, compaction is often either fragmentation-neutral or moving movable pages away from unmovable/reclaimable pages. As the actual watermarks are preserved, allow compaction to ignore the boost factor. 1-socket thpscale 4.20.0-rc6 4.20.0-rc6 finishscan-v1r4 noboost-v1r4 Amean fault-both-1 0.00 ( 0.00%) 0.00 * 0.00%* Amean fault-both-3 3849.90 ( 0.00%) 3753.53 ( 2.50%) Amean fault-both-5 5054.13 ( 0.00%) 5396.32 ( -6.77%) Amean fault-both-7 7061.77 ( 0.00%) 7393.46 ( -4.70%) Amean fault-both-12 11560.59 ( 0.00%) 12155.50 ( -5.15%) Amean fault-both-18 16120.15 ( 0.00%) 16445.96 ( -2.02%) Amean fault-both-24 19804.31 ( 0.00%) 20465.03 ( -3.34%) Amean fault-both-30 25018.73 ( 0.00%) 20813.54 * 16.81%* Amean fault-both-32 24380.19 ( 0.00%) 22384.02 ( 8.19%) The impact on the scan rates is a mixed bag because this patch is very sensitive to timing and whether the boost was active or not. However, detailed tracing indicated that failure of migration due to a premature ENOMEM triggered by watermark checks were eliminated. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> --- mm/page_alloc.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)