Re: [PATCH V2 0/6] Memory compaction efficiency improvements

From: Vlastimil Babka
Date: Thu Dec 12 2013 - 08:27:16 EST


On 12/12/2013 07:12 AM, Joonsoo Kim wrote:
On Wed, Dec 11, 2013 at 11:24:31AM +0100, Vlastimil Babka wrote:
Changelog since V1 (thanks to the reviewers!)
o Included "trace compaction being and end" patch in the series (mgorman)
o Changed variable names and comments in patches 2 and 5 (mgorman)
o More thorough measurements, based on v3.13-rc2

The broad goal of the series is to improve allocation success rates for huge
pages through memory compaction, while trying not to increase the compaction
overhead. The original objective was to reintroduce capturing of high-order
pages freed by the compaction, before they are split by concurrent activity.
However, several bugs and opportunities for simple improvements were found in
the current implementation, mostly through extra tracepoints (which are however
too ugly for now to be considered for sending).

The patches mostly deal with two mechanisms that reduce compaction overhead,
which is caching the progress of migrate and free scanners, and marking
pageblocks where isolation failed to be skipped during further scans.

Patch 1 (from mgorman) adds tracepoints that allow calculate time spent in
compaction and potentially debug scanner pfn values.

Patch 2 encapsulates the some functionality for handling deferred compactions
for better maintainability, without a functional change
type is not determined without being actually needed.

Patch 3 fixes a bug where cached scanner pfn's are sometimes reset only after
they have been read to initialize a compaction run.

Patch 4 fixes a bug where scanners meeting is sometimes not properly detected
and can lead to multiple compaction attempts quitting early without
doing any work.

Patch 5 improves the chances of sync compaction to process pageblocks that
async compaction has skipped due to being !MIGRATE_MOVABLE.

Patch 6 improves the chances of sync direct compaction to actually do anything
when called after async compaction fails during allocation slowpath.

The impact of patches were validated using mmtests's stress-highalloc benchmark
with mmtests's stress-highalloc benchmark on a x86_64 machine with 4GB memory.

Due to instability of the results (mostly related to the bugs fixed by patches
2 and 3), 10 iterations were performed, taking min,mean,max values for success
rates and mean values for time and vmstat-based metrics.

First, the default GFP_HIGHUSER_MOVABLE allocations were tested with the patches
stacked on top of v3.13-rc2. Patch 2 is OK to serve as baseline due to no
functional changes in 1 and 2. Comments below.

stress-highalloc
3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2
2-nothp 3-nothp 4-nothp 5-nothp 6-nothp
Success 1 Min 9.00 ( 0.00%) 10.00 (-11.11%) 43.00 (-377.78%) 43.00 (-377.78%) 33.00 (-266.67%)
Success 1 Mean 27.50 ( 0.00%) 25.30 ( 8.00%) 45.50 (-65.45%) 45.90 (-66.91%) 46.30 (-68.36%)
Success 1 Max 36.00 ( 0.00%) 36.00 ( 0.00%) 47.00 (-30.56%) 48.00 (-33.33%) 52.00 (-44.44%)
Success 2 Min 10.00 ( 0.00%) 8.00 ( 20.00%) 46.00 (-360.00%) 45.00 (-350.00%) 35.00 (-250.00%)
Success 2 Mean 26.40 ( 0.00%) 23.50 ( 10.98%) 47.30 (-79.17%) 47.60 (-80.30%) 48.10 (-82.20%)
Success 2 Max 34.00 ( 0.00%) 33.00 ( 2.94%) 48.00 (-41.18%) 50.00 (-47.06%) 54.00 (-58.82%)
Success 3 Min 65.00 ( 0.00%) 63.00 ( 3.08%) 85.00 (-30.77%) 84.00 (-29.23%) 85.00 (-30.77%)
Success 3 Mean 76.70 ( 0.00%) 70.50 ( 8.08%) 86.20 (-12.39%) 85.50 (-11.47%) 86.00 (-12.13%)
Success 3 Max 87.00 ( 0.00%) 86.00 ( 1.15%) 88.00 ( -1.15%) 87.00 ( 0.00%) 87.00 ( 0.00%)

3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2
2-nothp 3-nothp 4-nothp 5-nothp 6-nothp
User 6437.72 6459.76 5960.32 5974.55 6019.67
System 1049.65 1049.09 1029.32 1031.47 1032.31
Elapsed 1856.77 1874.48 1949.97 1994.22 1983.15

3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2
2-nothp 3-nothp 4-nothp 5-nothp 6-nothp
Minor Faults 253952267 254581900 250030122 250507333 250157829
Major Faults 420 407 506 530 530
Swap Ins 4 9 9 6 6
Swap Outs 398 375 345 346 333
Direct pages scanned 197538 189017 298574 287019 299063
Kswapd pages scanned 1809843 1801308 1846674 1873184 1861089
Kswapd pages reclaimed 1806972 1798684 1844219 1870509 1858622
Direct pages reclaimed 197227 188829 298380 286822 298835
Kswapd efficiency 99% 99% 99% 99% 99%
Kswapd velocity 953.382 970.449 952.243 934.569 922.286
Direct efficiency 99% 99% 99% 99% 99%
Direct velocity 104.058 101.832 153.961 143.200 148.205
Percentage direct scans 9% 9% 13% 13% 13%
Zone normal velocity 347.289 359.676 348.063 339.933 332.983
Zone dma32 velocity 710.151 712.605 758.140 737.835 737.507
Zone dma velocity 0.000 0.000 0.000 0.000 0.000
Page writes by reclaim 557.600 429.000 353.600 426.400 381.800
Page writes file 159 53 7 79 48
Page writes anon 398 375 345 346 333
Page reclaim immediate 825 644 411 575 420
Sector Reads 2781750 2769780 2878547 2939128 2910483
Sector Writes 12080843 12083351 12012892 12002132 12010745
Page rescued immediate 0 0 0 0 0
Slabs scanned 1575654 1545344 1778406 1786700 1794073
Direct inode steals 9657 10037 15795 14104 14645
Kswapd inode steals 46857 46335 50543 50716 51796
Kswapd skipped wait 0 0 0 0 0
THP fault alloc 97 91 81 71 77
THP collapse alloc 456 506 546 544 565
THP splits 6 5 5 4 4
THP fault fallback 0 1 0 0 0
THP collapse fail 14 14 12 13 12
Compaction stalls 1006 980 1537 1536 1548
Compaction success 303 284 562 559 578
Compaction failures 702 696 974 976 969
Page migrate success 1177325 1070077 3927538 3781870 3877057
Page migrate failure 0 0 0 0 0
Compaction pages isolated 2547248 2306457 8301218 8008500 8200674
Compaction migrate scanned 42290478 38832618 153961130 154143900 159141197
Compaction free scanned 89199429 79189151 356529027 351943166 356326727
Compaction cost 1566 1426 5312 5156 5294
NUMA PTE updates 0 0 0 0 0
NUMA hint faults 0 0 0 0 0
NUMA hint local faults 0 0 0 0 0
NUMA hint local percent 100 100 100 100 100
NUMA pages migrated 0 0 0 0 0
AutoNUMA cost 0 0 0 0 0


Observations:
- The "Success 3" line is allocation success rate with system idle (phases 1
and 2 are with background interference). I used to get stable values around
85% with vanilla 3.11. The lower min and mean values came with 3.12.
This was bisected to commit 81c0a2bb ("mm: page_alloc: fair zone allocator
policy") As explained in comment for patch 3, I don't think the commit is
wrong, but that it makes the effect of compaction bugs worse. From patch 3
onwards, the results are OK and match the 3.11 results.
- Patch 4 also clearly helps phases 1 and 2, and exceeds any results I've
seen with 3.11 (I didn't measure it that thoroughly then, but it was never
above 40%).
- Compaction cost and number of scanned pages is higher, especially due to
patch 4. However, keep in mind that patches 3 and 4 fix existing bugs in the
current design of compaction overhead mitigation, they do not change it.
If overhead is found unacceptable, then it should be decreased differently
(and consistently, not due to random conditions) than the current implementation
does. In contrast, patches 5 and 6 (which are not strictly bug fixes) do not
increase the overhead (but also not success rates). This might be a limitation
of the stress-highalloc benchmark as it's quite uniform.

Another set of results is when configuring stress-highalloc t allocate
with similar flags as THP uses:
(GFP_HIGHUSER_MOVABLE|__GFP_NOMEMALLOC|__GFP_NORETRY|__GFP_NO_KSWAPD)

stress-highalloc
3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2
2-thp 3-thp 4-thp 5-thp 6-thp
Success 1 Min 2.00 ( 0.00%) 7.00 (-250.00%) 18.00 (-800.00%) 19.00 (-850.00%) 26.00 (-1200.00%)
Success 1 Mean 19.20 ( 0.00%) 17.80 ( 7.29%) 29.20 (-52.08%) 29.90 (-55.73%) 32.80 (-70.83%)
Success 1 Max 27.00 ( 0.00%) 29.00 ( -7.41%) 35.00 (-29.63%) 36.00 (-33.33%) 37.00 (-37.04%)
Success 2 Min 3.00 ( 0.00%) 8.00 (-166.67%) 21.00 (-600.00%) 21.00 (-600.00%) 32.00 (-966.67%)
Success 2 Mean 19.30 ( 0.00%) 17.90 ( 7.25%) 32.20 (-66.84%) 32.60 (-68.91%) 35.70 (-84.97%)
Success 2 Max 27.00 ( 0.00%) 30.00 (-11.11%) 36.00 (-33.33%) 37.00 (-37.04%) 39.00 (-44.44%)
Success 3 Min 62.00 ( 0.00%) 62.00 ( 0.00%) 85.00 (-37.10%) 75.00 (-20.97%) 64.00 ( -3.23%)
Success 3 Mean 66.30 ( 0.00%) 65.50 ( 1.21%) 85.60 (-29.11%) 83.40 (-25.79%) 83.50 (-25.94%)
Success 3 Max 70.00 ( 0.00%) 69.00 ( 1.43%) 87.00 (-24.29%) 86.00 (-22.86%) 87.00 (-24.29%)

3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2
2-thp 3-thp 4-thp 5-thp 6-thp
User 6547.93 6475.85 6265.54 6289.46 6189.96
System 1053.42 1047.28 1043.23 1042.73 1038.73
Elapsed 1835.43 1821.96 1908.67 1912.74 1956.38

Hello, Vlastimil.

I have some questions related to your stat, not your patchset,
just for curiosity. :)

Are these results, "elapsed time" and "vmstat", for Success 3 line scenario?

No that's for the whole test which does the scenarios in succession.

If so, could you show me others?
I wonder why thp case consumes more system time rather than no-thp case.

Unfortunately these stats are not that useful as they don't distinguish the 3 phases and also include what the background load does. They are included just to show that nothing truly dramatic is happening.
So

And I found that elapsed time has no big difference between both cases,
roughly less than 2%. In this situation, do we get more benefits with
aggressive allocation like no-thp case?

Elapsed time suffers from the same problem, so it's again hard to say how relevant it actually is to the allocator workload and how much to background load. It seems that the more successful allocator is, the longer elapsed time (in both thp and nothp case). My guess is that less memory available for the background load makes it progress slower which affects the duration of the test as a whole.

I hope that in case of further compaction patches that would be potentially more intrusive to the its design (and not bugfixes and simple tweaks to the existing design as this series) I will have a more detailed breakdown of what time is spent where.

Thanks,
Vlastimil

Thanks.


3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2 3.13-rc2
2-thp 3-thp 4-thp 5-thp 6-thp
Minor Faults 256805673 253106328 253222299 249830289 251184418
Major Faults 395 375 423 434 448
Swap Ins 12 10 10 12 9
Swap Outs 530 537 487 455 415
Direct pages scanned 71859 86046 153244 152764 190713
Kswapd pages scanned 1900994 1870240 1898012 1892864 1880520
Kswapd pages reclaimed 1897814 1867428 1894939 1890125 1877924
Direct pages reclaimed 71766 85908 153167 152643 190600
Kswapd efficiency 99% 99% 99% 99% 99%
Kswapd velocity 1029.000 1067.782 1000.091 991.049 951.218
Direct efficiency 99% 99% 99% 99% 99%
Direct velocity 38.897 49.127 80.747 79.983 96.468
Percentage direct scans 3% 4% 7% 7% 9%
Zone normal velocity 351.377 372.494 348.910 341.689 335.310
Zone dma32 velocity 716.520 744.414 731.928 729.343 712.377
Zone dma velocity 0.000 0.000 0.000 0.000 0.000
Page writes by reclaim 669.300 604.000 545.700 538.900 429.900
Page writes file 138 66 58 83 14
Page writes anon 530 537 487 455 415
Page reclaim immediate 806 655 772 548 517
Sector Reads 2711956 2703239 2811602 2818248 2839459
Sector Writes 12163238 12018662 12038248 11954736 11994892
Page rescued immediate 0 0 0 0 0
Slabs scanned 1385088 1388364 1507968 1513292 1558656
Direct inode steals 1739 2564 4622 5496 6007
Kswapd inode steals 47461 46406 47804 48013 48466
Kswapd skipped wait 0 0 0 0 0
THP fault alloc 110 82 84 69 70
THP collapse alloc 445 482 467 462 539
THP splits 6 5 4 5 3
THP fault fallback 3 0 0 0 0
THP collapse fail 15 14 14 14 13
Compaction stalls 659 685 1033 1073 1111
Compaction success 222 225 410 427 456
Compaction failures 436 460 622 646 655
Page migrate success 446594 439978 1085640 1095062 1131716
Page migrate failure 0 0 0 0 0
Compaction pages isolated 1029475 1013490 2453074 2482698 2565400
Compaction migrate scanned 9955461 11344259 24375202 27978356 30494204
Compaction free scanned 27715272 28544654 80150615 82898631 85756132
Compaction cost 552 555 1344 1379 1436
NUMA PTE updates 0 0 0 0 0
NUMA hint faults 0 0 0 0 0
NUMA hint local faults 0 0 0 0 0
NUMA hint local percent 100 100 100 100 100
NUMA pages migrated 0 0 0 0 0
AutoNUMA cost 0 0 0 0 0

There are some differences from the previous results for THP-like allocations:
- Here, the bad result for unpatched kernel in phase 3 is much more consistent
to be between 65-70% and not related to the "regression" in 3.12. Still there is
the improvement from patch 4 onwards, which brings it on par with simple
GFP_HIGHUSER_MOVABLE allocations.
- Compaction costs have increased, but nowhere near as much as the non-THP case. Again,
the patches should be worth the gained determininsm.
- Patches 5 and 6 somewhat increase the number of migrate-scanned pages. This is most likely
due to __GFP_NO_KSWAPD flag, which means the cached pfn's and pageblock skip bits are not
reset by kswapd that often (at least in phase 3 where no concurrent activity would wake
up kswapd) and the patches thus help the sync-after-async compaction. It doesn't however
show that the sync compaction would help so much with success rates, which can be again
seen as a limitation of the benchmark scenario.



Mel Gorman (1):
mm: compaction: trace compaction begin and end

Vlastimil Babka (5):
mm: compaction: encapsulate defer reset logic
mm: compaction: reset cached scanner pfn's before reading them
mm: compaction: detect when scanners meet in isolate_freepages
mm: compaction: do not mark unmovable pageblocks as skipped in async
compaction
mm: compaction: reset scanner positions immediately when they meet

include/linux/compaction.h | 16 ++++++++++
include/trace/events/compaction.h | 42 +++++++++++++++++++++++++++
mm/compaction.c | 61 +++++++++++++++++++++++++++------------
mm/page_alloc.c | 5 +---
4 files changed, 102 insertions(+), 22 deletions(-)

--
1.8.4

--
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>

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/