diff options
author | Mel Gorman <mel@csn.ul.ie> | 2011-02-25 14:44:20 -0800 |
---|---|---|
committer | Linus Torvalds <torvalds@linux-foundation.org> | 2011-02-25 15:07:36 -0800 |
commit | 2876592f231d436c295b67726313f6f3cfb6e243 (patch) | |
tree | e53c6db2aed6e672481c31083287d79f32ad45f4 /mm/vmscan.c | |
parent | ac3c8304190ed0daaa2fb01ce2a069be5e2a52a7 (diff) | |
download | linux-2876592f231d436c295b67726313f6f3cfb6e243.tar.bz2 |
mm: vmscan: stop reclaim/compaction earlier due to insufficient progress if !__GFP_REPEAT
should_continue_reclaim() for reclaim/compaction allows scanning to
continue even if pages are not being reclaimed until the full list is
scanned. In terms of allocation success, this makes sense but potentially
it introduces unwanted latency for high-order allocations such as
transparent hugepages and network jumbo frames that would prefer to fail
the allocation attempt and fallback to order-0 pages. Worse, there is a
potential that the full LRU scan will clear all the young bits, distort
page aging information and potentially push pages into swap that would
have otherwise remained resident.
This patch will stop reclaim/compaction if no pages were reclaimed in the
last SWAP_CLUSTER_MAX pages that were considered. For allocations such as
hugetlbfs that use __GFP_REPEAT and have fewer fallback options, the full
LRU list may still be scanned.
Order-0 allocation should not be affected because RECLAIM_MODE_COMPACTION
is not set so the following avoids the gfp_mask being examined:
if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
return false;
A tool was developed based on ftrace that tracked the latency of
high-order allocations while transparent hugepage support was enabled and
three benchmarks were run. The "fix-infinite" figures are 2.6.38-rc4 with
Johannes's patch "vmscan: fix zone shrinking exit when scan work is done"
applied.
STREAM Highorder Allocation Latency Statistics
fix-infinite break-early
1 :: Count 10298 10229
1 :: Min 0.4560 0.4640
1 :: Mean 1.0589 1.0183
1 :: Max 14.5990 11.7510
1 :: Stddev 0.5208 0.4719
2 :: Count 2 1
2 :: Min 1.8610 3.7240
2 :: Mean 3.4325 3.7240
2 :: Max 5.0040 3.7240
2 :: Stddev 1.5715 0.0000
9 :: Count 111696 111694
9 :: Min 0.5230 0.4110
9 :: Mean 10.5831 10.5718
9 :: Max 38.4480 43.2900
9 :: Stddev 1.1147 1.1325
Mean time for order-1 allocations is reduced. order-2 looks increased but
with so few allocations, it's not particularly significant. THP mean
allocation latency is also reduced. That said, allocation time varies so
significantly that the reductions are within noise.
Max allocation time is reduced by a significant amount for low-order
allocations but reduced for THP allocations which presumably are now
breaking before reclaim has done enough work.
SysBench Highorder Allocation Latency Statistics
fix-infinite break-early
1 :: Count 15745 15677
1 :: Min 0.4250 0.4550
1 :: Mean 1.1023 1.0810
1 :: Max 14.4590 10.8220
1 :: Stddev 0.5117 0.5100
2 :: Count 1 1
2 :: Min 3.0040 2.1530
2 :: Mean 3.0040 2.1530
2 :: Max 3.0040 2.1530
2 :: Stddev 0.0000 0.0000
9 :: Count 2017 1931
9 :: Min 0.4980 0.7480
9 :: Mean 10.4717 10.3840
9 :: Max 24.9460 26.2500
9 :: Stddev 1.1726 1.1966
Again, mean time for order-1 allocations is reduced while order-2
allocations are too few to draw conclusions from. The mean time for THP
allocations is also slightly reduced albeit the reductions are within
varianes.
Once again, our maximum allocation time is significantly reduced for
low-order allocations and slightly increased for THP allocations.
Anon stream mmap reference Highorder Allocation Latency Statistics
1 :: Count 1376 1790
1 :: Min 0.4940 0.5010
1 :: Mean 1.0289 0.9732
1 :: Max 6.2670 4.2540
1 :: Stddev 0.4142 0.2785
2 :: Count 1 -
2 :: Min 1.9060 -
2 :: Mean 1.9060 -
2 :: Max 1.9060 -
2 :: Stddev 0.0000 -
9 :: Count 11266 11257
9 :: Min 0.4990 0.4940
9 :: Mean 27250.4669 24256.1919
9 :: Max 11439211.0000 6008885.0000
9 :: Stddev 226427.4624 186298.1430
This benchmark creates one thread per CPU which references an amount of
anonymous memory 1.5 times the size of physical RAM. This pounds swap
quite heavily and is intended to exercise THP a bit.
Mean allocation time for order-1 is reduced as before. It's also reduced
for THP allocations but the variations here are pretty massive due to
swap. As before, maximum allocation times are significantly reduced.
Overall, the patch reduces the mean and maximum allocation latencies for
the smaller high-order allocations. This was with Slab configured so it
would be expected to be more significant with Slub which uses these size
allocations more aggressively.
The mean allocation times for THP allocations are also slightly reduced.
The maximum latency was slightly increased as predicted by the comments
due to reclaim/compaction breaking early. However, workloads care more
about the latency of lower-order allocations than THP so it's an
acceptable trade-off.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'mm/vmscan.c')
-rw-r--r-- | mm/vmscan.c | 32 |
1 files changed, 22 insertions, 10 deletions
diff --git a/mm/vmscan.c b/mm/vmscan.c index 17497d0cd8b9..6771ea70bfe7 100644 --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -1841,16 +1841,28 @@ static inline bool should_continue_reclaim(struct zone *zone, if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION)) return false; - /* - * If we failed to reclaim and have scanned the full list, stop. - * NOTE: Checking just nr_reclaimed would exit reclaim/compaction far - * faster but obviously would be less likely to succeed - * allocation. If this is desirable, use GFP_REPEAT to decide - * if both reclaimed and scanned should be checked or just - * reclaimed - */ - if (!nr_reclaimed && !nr_scanned) - return false; + /* Consider stopping depending on scan and reclaim activity */ + if (sc->gfp_mask & __GFP_REPEAT) { + /* + * For __GFP_REPEAT allocations, stop reclaiming if the + * full LRU list has been scanned and we are still failing + * to reclaim pages. This full LRU scan is potentially + * expensive but a __GFP_REPEAT caller really wants to succeed + */ + if (!nr_reclaimed && !nr_scanned) + return false; + } else { + /* + * For non-__GFP_REPEAT allocations which can presumably + * fail without consequence, stop if we failed to reclaim + * any pages from the last SWAP_CLUSTER_MAX number of + * pages that were scanned. This will return to the + * caller faster at the risk reclaim/compaction and + * the resulting allocation attempt fails + */ + if (!nr_reclaimed) + return false; + } /* * If we have not reclaimed enough pages for compaction and the |