OpenCloudOS-Kernel/include/trace/events/vmscan.h

391 lines
9.0 KiB
C
Raw Normal View History

#undef TRACE_SYSTEM
#define TRACE_SYSTEM vmscan
#if !defined(_TRACE_VMSCAN_H) || defined(TRACE_HEADER_MULTI_READ)
#define _TRACE_VMSCAN_H
#include <linux/types.h>
#include <linux/tracepoint.h>
vmscan: implement swap token trace This is useful for observing swap token activity. example output: zsh-1845 [000] 598.962716: update_swap_token_priority: mm=ffff88015eaf7700 old_prio=1 new_prio=0 memtoy-1830 [001] 602.033900: update_swap_token_priority: mm=ffff880037a45880 old_prio=947 new_prio=949 memtoy-1830 [000] 602.041509: update_swap_token_priority: mm=ffff880037a45880 old_prio=949 new_prio=951 memtoy-1830 [000] 602.051959: update_swap_token_priority: mm=ffff880037a45880 old_prio=951 new_prio=953 memtoy-1830 [000] 602.052188: update_swap_token_priority: mm=ffff880037a45880 old_prio=953 new_prio=955 memtoy-1830 [001] 602.427184: put_swap_token: token_mm=ffff880037a45880 zsh-1789 [000] 602.427281: replace_swap_token: old_token_mm= (null) old_prio=0 new_token_mm=ffff88015eaf7018 new_prio=2 zsh-1789 [001] 602.433456: update_swap_token_priority: mm=ffff88015eaf7018 old_prio=2 new_prio=4 zsh-1789 [000] 602.437613: update_swap_token_priority: mm=ffff88015eaf7018 old_prio=4 new_prio=6 zsh-1789 [000] 602.443924: update_swap_token_priority: mm=ffff88015eaf7018 old_prio=6 new_prio=8 zsh-1789 [000] 602.451873: update_swap_token_priority: mm=ffff88015eaf7018 old_prio=8 new_prio=10 zsh-1789 [001] 602.462639: update_swap_token_priority: mm=ffff88015eaf7018 old_prio=10 new_prio=12 Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Acked-by: Rik van Riel<riel@redhat.com> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-06-16 06:08:14 +08:00
#include <linux/mm.h>
#include <linux/memcontrol.h>
#include <trace/events/gfpflags.h>
#define RECLAIM_WB_ANON 0x0001u
#define RECLAIM_WB_FILE 0x0002u
tracing, vmscan: add trace events for LRU list shrinking There have been numerous reports of stalls that pointed at the problem being somewhere in the VM. There are multiple roots to the problems which means dealing with any of the root problems in isolation is tricky to justify on their own and they would still need integration testing. This patch series puts together two different patch sets which in combination should tackle some of the root causes of latency problems being reported. Patch 1 adds a tracepoint for shrink_inactive_list. For this series, the most important results is being able to calculate the scanning/reclaim ratio as a measure of the amount of work being done by page reclaim. Patch 2 accounts for time spent in congestion_wait. Patches 3-6 were originally developed by Kosaki Motohiro but reworked for this series. It has been noted that lumpy reclaim is far too aggressive and trashes the system somewhat. As SLUB uses high-order allocations, a large cost incurred by lumpy reclaim will be noticeable. It was also reported during transparent hugepage support testing that lumpy reclaim was trashing the system and these patches should mitigate that problem without disabling lumpy reclaim. Patch 7 adds wait_iff_congested() and replaces some callers of congestion_wait(). wait_iff_congested() only sleeps if there is a BDI that is currently congested. Patch 8 notes that any BDI being congested is not necessarily a problem because there could be multiple BDIs of varying speeds and numberous zones. It attempts to track when a zone being reclaimed contains many pages backed by a congested BDI and if so, reclaimers wait on the congestion queue. I ran a number of tests with monitoring on X86, X86-64 and PPC64. Each machine had 3G of RAM and the CPUs were X86: Intel P4 2-core X86-64: AMD Phenom 4-core PPC64: PPC970MP Each used a single disk and the onboard IO controller. Dirty ratio was left at 20. I'm just going to report for X86-64 and PPC64 in a vague attempt to keep this report short. Four kernels were tested each based on v2.6.36-rc4 traceonly-v2r2: Patches 1 and 2 to instrument vmscan reclaims and congestion_wait lowlumpy-v2r3: Patches 1-6 to test if lumpy reclaim is better waitcongest-v2r3: Patches 1-7 to only wait on congestion waitwriteback-v2r4: Patches 1-8 to detect when a zone is congested nocongest-v1r5: Patches 1-3 for testing wait_iff_congestion nodirect-v1r5: Patches 1-10 to disable filesystem writeback for better IO The tests run were as follows kernbench compile-based benchmark. Smoke test performance sysbench OLTP read-only benchmark. Will be re-run in the future as read-write micro-mapped-file-stream This is a micro-benchmark from Johannes Weiner that accesses a large sparse-file through mmap(). It was configured to run in only single-CPU mode but can be indicative of how well page reclaim identifies suitable pages. stress-highalloc Tries to allocate huge pages under heavy load. kernbench, iozone and sysbench did not report any performance regression on any machine. sysbench did pressure the system lightly and there was reclaim activity but there were no difference of major interest between the kernels. X86-64 micro-mapped-file-stream traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 pgalloc_dma 1639.00 ( 0.00%) 667.00 (-145.73%) 1167.00 ( -40.45%) 578.00 (-183.56%) pgalloc_dma32 2842410.00 ( 0.00%) 2842626.00 ( 0.01%) 2843043.00 ( 0.02%) 2843014.00 ( 0.02%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 729.00 ( 0.00%) 85.00 (-757.65%) 609.00 ( -19.70%) 125.00 (-483.20%) pgsteal_dma32 2338721.00 ( 0.00%) 2447354.00 ( 4.44%) 2429536.00 ( 3.74%) 2436772.00 ( 4.02%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 1469.00 ( 0.00%) 532.00 (-176.13%) 1078.00 ( -36.27%) 220.00 (-567.73%) pgscan_kswapd_dma32 4597713.00 ( 0.00%) 4503597.00 ( -2.09%) 4295673.00 ( -7.03%) 3891686.00 ( -18.14%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 71.00 ( 0.00%) 134.00 ( 47.01%) 243.00 ( 70.78%) 352.00 ( 79.83%) pgscan_direct_dma32 305820.00 ( 0.00%) 280204.00 ( -9.14%) 600518.00 ( 49.07%) 957485.00 ( 68.06%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 16296.00 ( 0.00%) 21254.00 ( 23.33%) 18447.00 ( 11.66%) 20067.00 ( 18.79%) allocstall 443.00 ( 0.00%) 273.00 ( -62.27%) 513.00 ( 13.65%) 1568.00 ( 71.75%) These are based on the raw figures taken from /proc/vmstat. It's a rough measure of reclaim activity. Note that allocstall counts are higher because we are entering direct reclaim more often as a result of not sleeping in congestion. In itself, it's not necessarily a bad thing. It's easier to get a view of what happened from the vmscan tracepoint report. FTrace Reclaim Statistics: vmscan traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 Direct reclaims 443 273 513 1568 Direct reclaim pages scanned 305968 280402 600825 957933 Direct reclaim pages reclaimed 43503 19005 30327 117191 Direct reclaim write file async I/O 0 0 0 0 Direct reclaim write anon async I/O 0 3 4 12 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 187649 132338 191695 267701 Kswapd wakeups 3 1 4 1 Kswapd pages scanned 4599269 4454162 4296815 3891906 Kswapd pages reclaimed 2295947 2428434 2399818 2319706 Kswapd reclaim write file async I/O 1 0 1 1 Kswapd reclaim write anon async I/O 59 187 41 222 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4.34 2.52 6.63 2.96 Time kswapd awake (seconds) 11.15 10.25 11.01 10.19 Total pages scanned 4905237 4734564 4897640 4849839 Total pages reclaimed 2339450 2447439 2430145 2436897 %age total pages scanned/reclaimed 47.69% 51.69% 49.62% 50.25% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 29.23% 19.02% 38.48% 20.25% Percentage Time kswapd Awake 78.58% 78.85% 76.83% 79.86% What is interesting here for nocongest in particular is that while direct reclaim scans more pages, the overall number of pages scanned remains the same and the ratio of pages scanned to pages reclaimed is more or less the same. In other words, while we are sleeping less, reclaim is not doing more work and as direct reclaim and kswapd is awake for less time, it would appear to be doing less work. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 87 196 64 0 Direct time congest waited 4604ms 4732ms 5420ms 0ms Direct full congest waited 72 145 53 0 Direct number conditional waited 0 0 324 1315 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 20 10 15 7 KSwapd time congest waited 1264ms 536ms 884ms 284ms KSwapd full congest waited 10 4 6 2 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 8 seconds asleep in direct reclaim and no time at all asleep with the patches. MMTests Statistics: duration User/Sys Time Running Test (seconds) 10.51 10.73 10.6 11.66 Total Elapsed Time (seconds) 14.19 13.00 14.33 12.76 Overall, the tests completed faster. It is interesting to note that backing off further when a zone is congested and not just a BDI was more efficient overall. PPC64 micro-mapped-file-stream pgalloc_dma 3024660.00 ( 0.00%) 3027185.00 ( 0.08%) 3025845.00 ( 0.04%) 3026281.00 ( 0.05%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 2508073.00 ( 0.00%) 2565351.00 ( 2.23%) 2463577.00 ( -1.81%) 2532263.00 ( 0.96%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 4601307.00 ( 0.00%) 4128076.00 ( -11.46%) 3912317.00 ( -17.61%) 3377165.00 ( -36.25%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 629825.00 ( 0.00%) 971622.00 ( 35.18%) 1063938.00 ( 40.80%) 1711935.00 ( 63.21%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 27776.00 ( 0.00%) 20458.00 ( -35.77%) 18763.00 ( -48.04%) 18157.00 ( -52.98%) allocstall 977.00 ( 0.00%) 2751.00 ( 64.49%) 2098.00 ( 53.43%) 5136.00 ( 80.98%) Similar trends to x86-64. allocstalls are up but it's not necessarily bad. FTrace Reclaim Statistics: vmscan Direct reclaims 977 2709 2098 5136 Direct reclaim pages scanned 629825 963814 1063938 1711935 Direct reclaim pages reclaimed 75550 242538 150904 387647 Direct reclaim write file async I/O 0 0 0 2 Direct reclaim write anon async I/O 0 10 0 4 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 392119 1201712 571935 571921 Kswapd wakeups 3 2 3 3 Kswapd pages scanned 4601307 4128076 3912317 3377165 Kswapd pages reclaimed 2432523 2318797 2312673 2144616 Kswapd reclaim write file async I/O 20 1 1 1 Kswapd reclaim write anon async I/O 57 132 11 121 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 6.19 7.30 13.04 10.88 Time kswapd awake (seconds) 21.73 26.51 25.55 23.90 Total pages scanned 5231132 5091890 4976255 5089100 Total pages reclaimed 2508073 2561335 2463577 2532263 %age total pages scanned/reclaimed 47.95% 50.30% 49.51% 49.76% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 18.89% 20.65% 32.65% 27.65% Percentage Time kswapd Awake 72.39% 80.68% 78.21% 77.40% Again, a similar trend that the congestion_wait changes mean that direct reclaim scans more pages but the overall number of pages scanned while slightly reduced, are very similar. The ratio of scanning/reclaimed remains roughly similar. The downside is that kswapd and direct reclaim was awake longer and for a larger percentage of the overall workload. It's possible there were big differences in the amount of time spent reclaiming slab pages between the different kernels which is plausible considering that the micro tests runs after fsmark and sysbench. Trace Reclaim Statistics: congestion_wait Direct number congest waited 845 1312 104 0 Direct time congest waited 19416ms 26560ms 7544ms 0ms Direct full congest waited 745 1105 72 0 Direct number conditional waited 0 0 1322 2935 Direct time conditional waited 0ms 0ms 12ms 312ms Direct full conditional waited 0 0 0 3 KSwapd number congest waited 39 102 75 63 KSwapd time congest waited 2484ms 6760ms 5756ms 3716ms KSwapd full congest waited 20 48 46 25 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 20 seconds asleep in direct reclaim and only 312ms asleep with the patches. The time kswapd spent congest waited was also reduced by a large factor. MMTests Statistics: duration ser/Sys Time Running Test (seconds) 26.58 28.05 26.9 28.47 Total Elapsed Time (seconds) 30.02 32.86 32.67 30.88 With all patches applies, the completion times are very similar. X86-64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 82.00 ( 0.00%) 84.00 ( 2.00%) 85.00 ( 3.00%) 85.00 ( 3.00%) Pass 2 90.00 ( 0.00%) 87.00 (-3.00%) 88.00 (-2.00%) 89.00 (-1.00%) At Rest 92.00 ( 0.00%) 90.00 (-2.00%) 90.00 (-2.00%) 91.00 (-1.00%) Success figures across the board are broadly similar. traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 1045 944 886 887 Direct reclaim pages scanned 135091 119604 109382 101019 Direct reclaim pages reclaimed 88599 47535 47863 46671 Direct reclaim write file async I/O 494 283 465 280 Direct reclaim write anon async I/O 29357 13710 16656 13462 Direct reclaim write file sync I/O 154 2 2 3 Direct reclaim write anon sync I/O 14594 571 509 561 Wake kswapd requests 7491 933 872 892 Kswapd wakeups 814 778 731 780 Kswapd pages scanned 7290822 15341158 11916436 13703442 Kswapd pages reclaimed 3587336 3142496 3094392 3187151 Kswapd reclaim write file async I/O 91975 32317 28022 29628 Kswapd reclaim write anon async I/O 1992022 789307 829745 849769 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4588.93 2467.16 2495.41 2547.07 Time kswapd awake (seconds) 2497.66 1020.16 1098.06 1176.82 Total pages scanned 7425913 15460762 12025818 13804461 Total pages reclaimed 3675935 3190031 3142255 3233822 %age total pages scanned/reclaimed 49.50% 20.63% 26.13% 23.43% %age total pages scanned/written 28.66% 5.41% 7.28% 6.47% %age file pages scanned/written 1.25% 0.21% 0.24% 0.22% Percentage Time Spent Direct Reclaim 57.33% 42.15% 42.41% 42.99% Percentage Time kswapd Awake 43.56% 27.87% 29.76% 31.25% Scanned/reclaimed ratios again look good with big improvements in efficiency. The Scanned/written ratios also look much improved. With a better scanned/written ration, there is an expectation that IO would be more efficient and indeed, the time spent in direct reclaim is much reduced by the full series and kswapd spends a little less time awake. Overall, indications here are that allocations were happening much faster and this can be seen with a graph of the latency figures as the allocations were taking place http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-hydra-mean.ps FTrace Reclaim Statistics: congestion_wait Direct number congest waited 1333 204 169 4 Direct time congest waited 78896ms 8288ms 7260ms 200ms Direct full congest waited 756 92 69 2 Direct number conditional waited 0 0 26 186 Direct time conditional waited 0ms 0ms 0ms 2504ms Direct full conditional waited 0 0 0 25 KSwapd number congest waited 4 395 227 282 KSwapd time congest waited 384ms 25136ms 10508ms 18380ms KSwapd full congest waited 3 232 98 176 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 KSwapd full conditional waited 318 0 312 9 Overall, the time spent speeping is reduced. kswapd is still hitting congestion_wait() but that is because there are callers remaining where it wasn't clear in advance if they should be changed to wait_iff_congested() or not. Overall the sleep imes are reduced though - from 79ish seconds to about 19. MMTests Statistics: duration User/Sys Time Running Test (seconds) 3415.43 3386.65 3388.39 3377.5 Total Elapsed Time (seconds) 5733.48 3660.33 3689.41 3765.39 With the full series, the time to complete the tests are reduced by 30% PPC64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 17.00 ( 0.00%) 34.00 (17.00%) 38.00 (21.00%) 43.00 (26.00%) Pass 2 25.00 ( 0.00%) 37.00 (12.00%) 42.00 (17.00%) 46.00 (21.00%) At Rest 49.00 ( 0.00%) 43.00 (-6.00%) 45.00 (-4.00%) 51.00 ( 2.00%) Success rates there are *way* up particularly considering that the 16MB huge pages on PPC64 mean that it's always much harder to allocate them. FTrace Reclaim Statistics: vmscan stress-highalloc stress-highalloc stress-highalloc stress-highalloc traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 499 505 564 509 Direct reclaim pages scanned 223478 41898 51818 45605 Direct reclaim pages reclaimed 137730 21148 27161 23455 Direct reclaim write file async I/O 399 136 162 136 Direct reclaim write anon async I/O 46977 2865 4686 3998 Direct reclaim write file sync I/O 29 0 1 3 Direct reclaim write anon sync I/O 31023 159 237 239 Wake kswapd requests 420 351 360 326 Kswapd wakeups 185 294 249 277 Kswapd pages scanned 15703488 16392500 17821724 17598737 Kswapd pages reclaimed 5808466 2908858 3139386 3145435 Kswapd reclaim write file async I/O 159938 18400 18717 13473 Kswapd reclaim write anon async I/O 3467554 228957 322799 234278 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 9665.35 1707.81 2374.32 1871.23 Time kswapd awake (seconds) 9401.21 1367.86 1951.75 1328.88 Total pages scanned 15926966 16434398 17873542 17644342 Total pages reclaimed 5946196 2930006 3166547 3168890 %age total pages scanned/reclaimed 37.33% 17.83% 17.72% 17.96% %age total pages scanned/written 23.27% 1.52% 1.94% 1.43% %age file pages scanned/written 1.01% 0.11% 0.11% 0.08% Percentage Time Spent Direct Reclaim 44.55% 35.10% 41.42% 36.91% Percentage Time kswapd Awake 86.71% 43.58% 52.67% 41.14% While the scanning rates are slightly up, the scanned/reclaimed and scanned/written figures are much improved. The time spent in direct reclaim and with kswapd are massively reduced, mostly by the lowlumpy patches. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 725 303 126 3 Direct time congest waited 45524ms 9180ms 5936ms 300ms Direct full congest waited 487 190 52 3 Direct number conditional waited 0 0 200 301 Direct time conditional waited 0ms 0ms 0ms 1904ms Direct full conditional waited 0 0 0 19 KSwapd number congest waited 0 2 23 4 KSwapd time congest waited 0ms 200ms 420ms 404ms KSwapd full congest waited 0 2 2 4 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Not as dramatic a story here but the time spent asleep is reduced and we can still see what wait_iff_congested is going to sleep when necessary. MMTests Statistics: duration User/Sys Time Running Test (seconds) 12028.09 3157.17 3357.79 3199.16 Total Elapsed Time (seconds) 10842.07 3138.72 3705.54 3229.85 The time to complete this test goes way down. With the full series, we are allocating over twice the number of huge pages in 30% of the time and there is a corresponding impact on the allocation latency graph available at. http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-powyah-mean.ps This patch: Add a trace event for shrink_inactive_list() and updates the sample postprocessing script appropriately. It can be used to determine how many pages were reclaimed and for non-lumpy reclaim where exactly the pages were reclaimed from. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-27 05:21:40 +08:00
#define RECLAIM_WB_MIXED 0x0010u
#define RECLAIM_WB_SYNC 0x0004u /* Unused, all reclaim async */
#define RECLAIM_WB_ASYNC 0x0008u
#define show_reclaim_flags(flags) \
(flags) ? __print_flags(flags, "|", \
{RECLAIM_WB_ANON, "RECLAIM_WB_ANON"}, \
{RECLAIM_WB_FILE, "RECLAIM_WB_FILE"}, \
tracing, vmscan: add trace events for LRU list shrinking There have been numerous reports of stalls that pointed at the problem being somewhere in the VM. There are multiple roots to the problems which means dealing with any of the root problems in isolation is tricky to justify on their own and they would still need integration testing. This patch series puts together two different patch sets which in combination should tackle some of the root causes of latency problems being reported. Patch 1 adds a tracepoint for shrink_inactive_list. For this series, the most important results is being able to calculate the scanning/reclaim ratio as a measure of the amount of work being done by page reclaim. Patch 2 accounts for time spent in congestion_wait. Patches 3-6 were originally developed by Kosaki Motohiro but reworked for this series. It has been noted that lumpy reclaim is far too aggressive and trashes the system somewhat. As SLUB uses high-order allocations, a large cost incurred by lumpy reclaim will be noticeable. It was also reported during transparent hugepage support testing that lumpy reclaim was trashing the system and these patches should mitigate that problem without disabling lumpy reclaim. Patch 7 adds wait_iff_congested() and replaces some callers of congestion_wait(). wait_iff_congested() only sleeps if there is a BDI that is currently congested. Patch 8 notes that any BDI being congested is not necessarily a problem because there could be multiple BDIs of varying speeds and numberous zones. It attempts to track when a zone being reclaimed contains many pages backed by a congested BDI and if so, reclaimers wait on the congestion queue. I ran a number of tests with monitoring on X86, X86-64 and PPC64. Each machine had 3G of RAM and the CPUs were X86: Intel P4 2-core X86-64: AMD Phenom 4-core PPC64: PPC970MP Each used a single disk and the onboard IO controller. Dirty ratio was left at 20. I'm just going to report for X86-64 and PPC64 in a vague attempt to keep this report short. Four kernels were tested each based on v2.6.36-rc4 traceonly-v2r2: Patches 1 and 2 to instrument vmscan reclaims and congestion_wait lowlumpy-v2r3: Patches 1-6 to test if lumpy reclaim is better waitcongest-v2r3: Patches 1-7 to only wait on congestion waitwriteback-v2r4: Patches 1-8 to detect when a zone is congested nocongest-v1r5: Patches 1-3 for testing wait_iff_congestion nodirect-v1r5: Patches 1-10 to disable filesystem writeback for better IO The tests run were as follows kernbench compile-based benchmark. Smoke test performance sysbench OLTP read-only benchmark. Will be re-run in the future as read-write micro-mapped-file-stream This is a micro-benchmark from Johannes Weiner that accesses a large sparse-file through mmap(). It was configured to run in only single-CPU mode but can be indicative of how well page reclaim identifies suitable pages. stress-highalloc Tries to allocate huge pages under heavy load. kernbench, iozone and sysbench did not report any performance regression on any machine. sysbench did pressure the system lightly and there was reclaim activity but there were no difference of major interest between the kernels. X86-64 micro-mapped-file-stream traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 pgalloc_dma 1639.00 ( 0.00%) 667.00 (-145.73%) 1167.00 ( -40.45%) 578.00 (-183.56%) pgalloc_dma32 2842410.00 ( 0.00%) 2842626.00 ( 0.01%) 2843043.00 ( 0.02%) 2843014.00 ( 0.02%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 729.00 ( 0.00%) 85.00 (-757.65%) 609.00 ( -19.70%) 125.00 (-483.20%) pgsteal_dma32 2338721.00 ( 0.00%) 2447354.00 ( 4.44%) 2429536.00 ( 3.74%) 2436772.00 ( 4.02%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 1469.00 ( 0.00%) 532.00 (-176.13%) 1078.00 ( -36.27%) 220.00 (-567.73%) pgscan_kswapd_dma32 4597713.00 ( 0.00%) 4503597.00 ( -2.09%) 4295673.00 ( -7.03%) 3891686.00 ( -18.14%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 71.00 ( 0.00%) 134.00 ( 47.01%) 243.00 ( 70.78%) 352.00 ( 79.83%) pgscan_direct_dma32 305820.00 ( 0.00%) 280204.00 ( -9.14%) 600518.00 ( 49.07%) 957485.00 ( 68.06%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 16296.00 ( 0.00%) 21254.00 ( 23.33%) 18447.00 ( 11.66%) 20067.00 ( 18.79%) allocstall 443.00 ( 0.00%) 273.00 ( -62.27%) 513.00 ( 13.65%) 1568.00 ( 71.75%) These are based on the raw figures taken from /proc/vmstat. It's a rough measure of reclaim activity. Note that allocstall counts are higher because we are entering direct reclaim more often as a result of not sleeping in congestion. In itself, it's not necessarily a bad thing. It's easier to get a view of what happened from the vmscan tracepoint report. FTrace Reclaim Statistics: vmscan traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 Direct reclaims 443 273 513 1568 Direct reclaim pages scanned 305968 280402 600825 957933 Direct reclaim pages reclaimed 43503 19005 30327 117191 Direct reclaim write file async I/O 0 0 0 0 Direct reclaim write anon async I/O 0 3 4 12 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 187649 132338 191695 267701 Kswapd wakeups 3 1 4 1 Kswapd pages scanned 4599269 4454162 4296815 3891906 Kswapd pages reclaimed 2295947 2428434 2399818 2319706 Kswapd reclaim write file async I/O 1 0 1 1 Kswapd reclaim write anon async I/O 59 187 41 222 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4.34 2.52 6.63 2.96 Time kswapd awake (seconds) 11.15 10.25 11.01 10.19 Total pages scanned 4905237 4734564 4897640 4849839 Total pages reclaimed 2339450 2447439 2430145 2436897 %age total pages scanned/reclaimed 47.69% 51.69% 49.62% 50.25% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 29.23% 19.02% 38.48% 20.25% Percentage Time kswapd Awake 78.58% 78.85% 76.83% 79.86% What is interesting here for nocongest in particular is that while direct reclaim scans more pages, the overall number of pages scanned remains the same and the ratio of pages scanned to pages reclaimed is more or less the same. In other words, while we are sleeping less, reclaim is not doing more work and as direct reclaim and kswapd is awake for less time, it would appear to be doing less work. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 87 196 64 0 Direct time congest waited 4604ms 4732ms 5420ms 0ms Direct full congest waited 72 145 53 0 Direct number conditional waited 0 0 324 1315 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 20 10 15 7 KSwapd time congest waited 1264ms 536ms 884ms 284ms KSwapd full congest waited 10 4 6 2 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 8 seconds asleep in direct reclaim and no time at all asleep with the patches. MMTests Statistics: duration User/Sys Time Running Test (seconds) 10.51 10.73 10.6 11.66 Total Elapsed Time (seconds) 14.19 13.00 14.33 12.76 Overall, the tests completed faster. It is interesting to note that backing off further when a zone is congested and not just a BDI was more efficient overall. PPC64 micro-mapped-file-stream pgalloc_dma 3024660.00 ( 0.00%) 3027185.00 ( 0.08%) 3025845.00 ( 0.04%) 3026281.00 ( 0.05%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 2508073.00 ( 0.00%) 2565351.00 ( 2.23%) 2463577.00 ( -1.81%) 2532263.00 ( 0.96%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 4601307.00 ( 0.00%) 4128076.00 ( -11.46%) 3912317.00 ( -17.61%) 3377165.00 ( -36.25%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 629825.00 ( 0.00%) 971622.00 ( 35.18%) 1063938.00 ( 40.80%) 1711935.00 ( 63.21%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 27776.00 ( 0.00%) 20458.00 ( -35.77%) 18763.00 ( -48.04%) 18157.00 ( -52.98%) allocstall 977.00 ( 0.00%) 2751.00 ( 64.49%) 2098.00 ( 53.43%) 5136.00 ( 80.98%) Similar trends to x86-64. allocstalls are up but it's not necessarily bad. FTrace Reclaim Statistics: vmscan Direct reclaims 977 2709 2098 5136 Direct reclaim pages scanned 629825 963814 1063938 1711935 Direct reclaim pages reclaimed 75550 242538 150904 387647 Direct reclaim write file async I/O 0 0 0 2 Direct reclaim write anon async I/O 0 10 0 4 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 392119 1201712 571935 571921 Kswapd wakeups 3 2 3 3 Kswapd pages scanned 4601307 4128076 3912317 3377165 Kswapd pages reclaimed 2432523 2318797 2312673 2144616 Kswapd reclaim write file async I/O 20 1 1 1 Kswapd reclaim write anon async I/O 57 132 11 121 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 6.19 7.30 13.04 10.88 Time kswapd awake (seconds) 21.73 26.51 25.55 23.90 Total pages scanned 5231132 5091890 4976255 5089100 Total pages reclaimed 2508073 2561335 2463577 2532263 %age total pages scanned/reclaimed 47.95% 50.30% 49.51% 49.76% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 18.89% 20.65% 32.65% 27.65% Percentage Time kswapd Awake 72.39% 80.68% 78.21% 77.40% Again, a similar trend that the congestion_wait changes mean that direct reclaim scans more pages but the overall number of pages scanned while slightly reduced, are very similar. The ratio of scanning/reclaimed remains roughly similar. The downside is that kswapd and direct reclaim was awake longer and for a larger percentage of the overall workload. It's possible there were big differences in the amount of time spent reclaiming slab pages between the different kernels which is plausible considering that the micro tests runs after fsmark and sysbench. Trace Reclaim Statistics: congestion_wait Direct number congest waited 845 1312 104 0 Direct time congest waited 19416ms 26560ms 7544ms 0ms Direct full congest waited 745 1105 72 0 Direct number conditional waited 0 0 1322 2935 Direct time conditional waited 0ms 0ms 12ms 312ms Direct full conditional waited 0 0 0 3 KSwapd number congest waited 39 102 75 63 KSwapd time congest waited 2484ms 6760ms 5756ms 3716ms KSwapd full congest waited 20 48 46 25 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 20 seconds asleep in direct reclaim and only 312ms asleep with the patches. The time kswapd spent congest waited was also reduced by a large factor. MMTests Statistics: duration ser/Sys Time Running Test (seconds) 26.58 28.05 26.9 28.47 Total Elapsed Time (seconds) 30.02 32.86 32.67 30.88 With all patches applies, the completion times are very similar. X86-64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 82.00 ( 0.00%) 84.00 ( 2.00%) 85.00 ( 3.00%) 85.00 ( 3.00%) Pass 2 90.00 ( 0.00%) 87.00 (-3.00%) 88.00 (-2.00%) 89.00 (-1.00%) At Rest 92.00 ( 0.00%) 90.00 (-2.00%) 90.00 (-2.00%) 91.00 (-1.00%) Success figures across the board are broadly similar. traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 1045 944 886 887 Direct reclaim pages scanned 135091 119604 109382 101019 Direct reclaim pages reclaimed 88599 47535 47863 46671 Direct reclaim write file async I/O 494 283 465 280 Direct reclaim write anon async I/O 29357 13710 16656 13462 Direct reclaim write file sync I/O 154 2 2 3 Direct reclaim write anon sync I/O 14594 571 509 561 Wake kswapd requests 7491 933 872 892 Kswapd wakeups 814 778 731 780 Kswapd pages scanned 7290822 15341158 11916436 13703442 Kswapd pages reclaimed 3587336 3142496 3094392 3187151 Kswapd reclaim write file async I/O 91975 32317 28022 29628 Kswapd reclaim write anon async I/O 1992022 789307 829745 849769 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4588.93 2467.16 2495.41 2547.07 Time kswapd awake (seconds) 2497.66 1020.16 1098.06 1176.82 Total pages scanned 7425913 15460762 12025818 13804461 Total pages reclaimed 3675935 3190031 3142255 3233822 %age total pages scanned/reclaimed 49.50% 20.63% 26.13% 23.43% %age total pages scanned/written 28.66% 5.41% 7.28% 6.47% %age file pages scanned/written 1.25% 0.21% 0.24% 0.22% Percentage Time Spent Direct Reclaim 57.33% 42.15% 42.41% 42.99% Percentage Time kswapd Awake 43.56% 27.87% 29.76% 31.25% Scanned/reclaimed ratios again look good with big improvements in efficiency. The Scanned/written ratios also look much improved. With a better scanned/written ration, there is an expectation that IO would be more efficient and indeed, the time spent in direct reclaim is much reduced by the full series and kswapd spends a little less time awake. Overall, indications here are that allocations were happening much faster and this can be seen with a graph of the latency figures as the allocations were taking place http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-hydra-mean.ps FTrace Reclaim Statistics: congestion_wait Direct number congest waited 1333 204 169 4 Direct time congest waited 78896ms 8288ms 7260ms 200ms Direct full congest waited 756 92 69 2 Direct number conditional waited 0 0 26 186 Direct time conditional waited 0ms 0ms 0ms 2504ms Direct full conditional waited 0 0 0 25 KSwapd number congest waited 4 395 227 282 KSwapd time congest waited 384ms 25136ms 10508ms 18380ms KSwapd full congest waited 3 232 98 176 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 KSwapd full conditional waited 318 0 312 9 Overall, the time spent speeping is reduced. kswapd is still hitting congestion_wait() but that is because there are callers remaining where it wasn't clear in advance if they should be changed to wait_iff_congested() or not. Overall the sleep imes are reduced though - from 79ish seconds to about 19. MMTests Statistics: duration User/Sys Time Running Test (seconds) 3415.43 3386.65 3388.39 3377.5 Total Elapsed Time (seconds) 5733.48 3660.33 3689.41 3765.39 With the full series, the time to complete the tests are reduced by 30% PPC64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 17.00 ( 0.00%) 34.00 (17.00%) 38.00 (21.00%) 43.00 (26.00%) Pass 2 25.00 ( 0.00%) 37.00 (12.00%) 42.00 (17.00%) 46.00 (21.00%) At Rest 49.00 ( 0.00%) 43.00 (-6.00%) 45.00 (-4.00%) 51.00 ( 2.00%) Success rates there are *way* up particularly considering that the 16MB huge pages on PPC64 mean that it's always much harder to allocate them. FTrace Reclaim Statistics: vmscan stress-highalloc stress-highalloc stress-highalloc stress-highalloc traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 499 505 564 509 Direct reclaim pages scanned 223478 41898 51818 45605 Direct reclaim pages reclaimed 137730 21148 27161 23455 Direct reclaim write file async I/O 399 136 162 136 Direct reclaim write anon async I/O 46977 2865 4686 3998 Direct reclaim write file sync I/O 29 0 1 3 Direct reclaim write anon sync I/O 31023 159 237 239 Wake kswapd requests 420 351 360 326 Kswapd wakeups 185 294 249 277 Kswapd pages scanned 15703488 16392500 17821724 17598737 Kswapd pages reclaimed 5808466 2908858 3139386 3145435 Kswapd reclaim write file async I/O 159938 18400 18717 13473 Kswapd reclaim write anon async I/O 3467554 228957 322799 234278 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 9665.35 1707.81 2374.32 1871.23 Time kswapd awake (seconds) 9401.21 1367.86 1951.75 1328.88 Total pages scanned 15926966 16434398 17873542 17644342 Total pages reclaimed 5946196 2930006 3166547 3168890 %age total pages scanned/reclaimed 37.33% 17.83% 17.72% 17.96% %age total pages scanned/written 23.27% 1.52% 1.94% 1.43% %age file pages scanned/written 1.01% 0.11% 0.11% 0.08% Percentage Time Spent Direct Reclaim 44.55% 35.10% 41.42% 36.91% Percentage Time kswapd Awake 86.71% 43.58% 52.67% 41.14% While the scanning rates are slightly up, the scanned/reclaimed and scanned/written figures are much improved. The time spent in direct reclaim and with kswapd are massively reduced, mostly by the lowlumpy patches. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 725 303 126 3 Direct time congest waited 45524ms 9180ms 5936ms 300ms Direct full congest waited 487 190 52 3 Direct number conditional waited 0 0 200 301 Direct time conditional waited 0ms 0ms 0ms 1904ms Direct full conditional waited 0 0 0 19 KSwapd number congest waited 0 2 23 4 KSwapd time congest waited 0ms 200ms 420ms 404ms KSwapd full congest waited 0 2 2 4 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Not as dramatic a story here but the time spent asleep is reduced and we can still see what wait_iff_congested is going to sleep when necessary. MMTests Statistics: duration User/Sys Time Running Test (seconds) 12028.09 3157.17 3357.79 3199.16 Total Elapsed Time (seconds) 10842.07 3138.72 3705.54 3229.85 The time to complete this test goes way down. With the full series, we are allocating over twice the number of huge pages in 30% of the time and there is a corresponding impact on the allocation latency graph available at. http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-powyah-mean.ps This patch: Add a trace event for shrink_inactive_list() and updates the sample postprocessing script appropriately. It can be used to determine how many pages were reclaimed and for non-lumpy reclaim where exactly the pages were reclaimed from. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-27 05:21:40 +08:00
{RECLAIM_WB_MIXED, "RECLAIM_WB_MIXED"}, \
{RECLAIM_WB_SYNC, "RECLAIM_WB_SYNC"}, \
{RECLAIM_WB_ASYNC, "RECLAIM_WB_ASYNC"} \
) : "RECLAIM_WB_NONE"
#define trace_reclaim_flags(page) ( \
(page_is_file_cache(page) ? RECLAIM_WB_FILE : RECLAIM_WB_ANON) | \
(RECLAIM_WB_ASYNC) \
)
#define trace_shrink_flags(file) \
( \
(file ? RECLAIM_WB_FILE : RECLAIM_WB_ANON) | \
(RECLAIM_WB_ASYNC) \
tracing, vmscan: add trace events for LRU list shrinking There have been numerous reports of stalls that pointed at the problem being somewhere in the VM. There are multiple roots to the problems which means dealing with any of the root problems in isolation is tricky to justify on their own and they would still need integration testing. This patch series puts together two different patch sets which in combination should tackle some of the root causes of latency problems being reported. Patch 1 adds a tracepoint for shrink_inactive_list. For this series, the most important results is being able to calculate the scanning/reclaim ratio as a measure of the amount of work being done by page reclaim. Patch 2 accounts for time spent in congestion_wait. Patches 3-6 were originally developed by Kosaki Motohiro but reworked for this series. It has been noted that lumpy reclaim is far too aggressive and trashes the system somewhat. As SLUB uses high-order allocations, a large cost incurred by lumpy reclaim will be noticeable. It was also reported during transparent hugepage support testing that lumpy reclaim was trashing the system and these patches should mitigate that problem without disabling lumpy reclaim. Patch 7 adds wait_iff_congested() and replaces some callers of congestion_wait(). wait_iff_congested() only sleeps if there is a BDI that is currently congested. Patch 8 notes that any BDI being congested is not necessarily a problem because there could be multiple BDIs of varying speeds and numberous zones. It attempts to track when a zone being reclaimed contains many pages backed by a congested BDI and if so, reclaimers wait on the congestion queue. I ran a number of tests with monitoring on X86, X86-64 and PPC64. Each machine had 3G of RAM and the CPUs were X86: Intel P4 2-core X86-64: AMD Phenom 4-core PPC64: PPC970MP Each used a single disk and the onboard IO controller. Dirty ratio was left at 20. I'm just going to report for X86-64 and PPC64 in a vague attempt to keep this report short. Four kernels were tested each based on v2.6.36-rc4 traceonly-v2r2: Patches 1 and 2 to instrument vmscan reclaims and congestion_wait lowlumpy-v2r3: Patches 1-6 to test if lumpy reclaim is better waitcongest-v2r3: Patches 1-7 to only wait on congestion waitwriteback-v2r4: Patches 1-8 to detect when a zone is congested nocongest-v1r5: Patches 1-3 for testing wait_iff_congestion nodirect-v1r5: Patches 1-10 to disable filesystem writeback for better IO The tests run were as follows kernbench compile-based benchmark. Smoke test performance sysbench OLTP read-only benchmark. Will be re-run in the future as read-write micro-mapped-file-stream This is a micro-benchmark from Johannes Weiner that accesses a large sparse-file through mmap(). It was configured to run in only single-CPU mode but can be indicative of how well page reclaim identifies suitable pages. stress-highalloc Tries to allocate huge pages under heavy load. kernbench, iozone and sysbench did not report any performance regression on any machine. sysbench did pressure the system lightly and there was reclaim activity but there were no difference of major interest between the kernels. X86-64 micro-mapped-file-stream traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 pgalloc_dma 1639.00 ( 0.00%) 667.00 (-145.73%) 1167.00 ( -40.45%) 578.00 (-183.56%) pgalloc_dma32 2842410.00 ( 0.00%) 2842626.00 ( 0.01%) 2843043.00 ( 0.02%) 2843014.00 ( 0.02%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 729.00 ( 0.00%) 85.00 (-757.65%) 609.00 ( -19.70%) 125.00 (-483.20%) pgsteal_dma32 2338721.00 ( 0.00%) 2447354.00 ( 4.44%) 2429536.00 ( 3.74%) 2436772.00 ( 4.02%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 1469.00 ( 0.00%) 532.00 (-176.13%) 1078.00 ( -36.27%) 220.00 (-567.73%) pgscan_kswapd_dma32 4597713.00 ( 0.00%) 4503597.00 ( -2.09%) 4295673.00 ( -7.03%) 3891686.00 ( -18.14%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 71.00 ( 0.00%) 134.00 ( 47.01%) 243.00 ( 70.78%) 352.00 ( 79.83%) pgscan_direct_dma32 305820.00 ( 0.00%) 280204.00 ( -9.14%) 600518.00 ( 49.07%) 957485.00 ( 68.06%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 16296.00 ( 0.00%) 21254.00 ( 23.33%) 18447.00 ( 11.66%) 20067.00 ( 18.79%) allocstall 443.00 ( 0.00%) 273.00 ( -62.27%) 513.00 ( 13.65%) 1568.00 ( 71.75%) These are based on the raw figures taken from /proc/vmstat. It's a rough measure of reclaim activity. Note that allocstall counts are higher because we are entering direct reclaim more often as a result of not sleeping in congestion. In itself, it's not necessarily a bad thing. It's easier to get a view of what happened from the vmscan tracepoint report. FTrace Reclaim Statistics: vmscan traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 Direct reclaims 443 273 513 1568 Direct reclaim pages scanned 305968 280402 600825 957933 Direct reclaim pages reclaimed 43503 19005 30327 117191 Direct reclaim write file async I/O 0 0 0 0 Direct reclaim write anon async I/O 0 3 4 12 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 187649 132338 191695 267701 Kswapd wakeups 3 1 4 1 Kswapd pages scanned 4599269 4454162 4296815 3891906 Kswapd pages reclaimed 2295947 2428434 2399818 2319706 Kswapd reclaim write file async I/O 1 0 1 1 Kswapd reclaim write anon async I/O 59 187 41 222 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4.34 2.52 6.63 2.96 Time kswapd awake (seconds) 11.15 10.25 11.01 10.19 Total pages scanned 4905237 4734564 4897640 4849839 Total pages reclaimed 2339450 2447439 2430145 2436897 %age total pages scanned/reclaimed 47.69% 51.69% 49.62% 50.25% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 29.23% 19.02% 38.48% 20.25% Percentage Time kswapd Awake 78.58% 78.85% 76.83% 79.86% What is interesting here for nocongest in particular is that while direct reclaim scans more pages, the overall number of pages scanned remains the same and the ratio of pages scanned to pages reclaimed is more or less the same. In other words, while we are sleeping less, reclaim is not doing more work and as direct reclaim and kswapd is awake for less time, it would appear to be doing less work. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 87 196 64 0 Direct time congest waited 4604ms 4732ms 5420ms 0ms Direct full congest waited 72 145 53 0 Direct number conditional waited 0 0 324 1315 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 20 10 15 7 KSwapd time congest waited 1264ms 536ms 884ms 284ms KSwapd full congest waited 10 4 6 2 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 8 seconds asleep in direct reclaim and no time at all asleep with the patches. MMTests Statistics: duration User/Sys Time Running Test (seconds) 10.51 10.73 10.6 11.66 Total Elapsed Time (seconds) 14.19 13.00 14.33 12.76 Overall, the tests completed faster. It is interesting to note that backing off further when a zone is congested and not just a BDI was more efficient overall. PPC64 micro-mapped-file-stream pgalloc_dma 3024660.00 ( 0.00%) 3027185.00 ( 0.08%) 3025845.00 ( 0.04%) 3026281.00 ( 0.05%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 2508073.00 ( 0.00%) 2565351.00 ( 2.23%) 2463577.00 ( -1.81%) 2532263.00 ( 0.96%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 4601307.00 ( 0.00%) 4128076.00 ( -11.46%) 3912317.00 ( -17.61%) 3377165.00 ( -36.25%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 629825.00 ( 0.00%) 971622.00 ( 35.18%) 1063938.00 ( 40.80%) 1711935.00 ( 63.21%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 27776.00 ( 0.00%) 20458.00 ( -35.77%) 18763.00 ( -48.04%) 18157.00 ( -52.98%) allocstall 977.00 ( 0.00%) 2751.00 ( 64.49%) 2098.00 ( 53.43%) 5136.00 ( 80.98%) Similar trends to x86-64. allocstalls are up but it's not necessarily bad. FTrace Reclaim Statistics: vmscan Direct reclaims 977 2709 2098 5136 Direct reclaim pages scanned 629825 963814 1063938 1711935 Direct reclaim pages reclaimed 75550 242538 150904 387647 Direct reclaim write file async I/O 0 0 0 2 Direct reclaim write anon async I/O 0 10 0 4 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 392119 1201712 571935 571921 Kswapd wakeups 3 2 3 3 Kswapd pages scanned 4601307 4128076 3912317 3377165 Kswapd pages reclaimed 2432523 2318797 2312673 2144616 Kswapd reclaim write file async I/O 20 1 1 1 Kswapd reclaim write anon async I/O 57 132 11 121 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 6.19 7.30 13.04 10.88 Time kswapd awake (seconds) 21.73 26.51 25.55 23.90 Total pages scanned 5231132 5091890 4976255 5089100 Total pages reclaimed 2508073 2561335 2463577 2532263 %age total pages scanned/reclaimed 47.95% 50.30% 49.51% 49.76% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 18.89% 20.65% 32.65% 27.65% Percentage Time kswapd Awake 72.39% 80.68% 78.21% 77.40% Again, a similar trend that the congestion_wait changes mean that direct reclaim scans more pages but the overall number of pages scanned while slightly reduced, are very similar. The ratio of scanning/reclaimed remains roughly similar. The downside is that kswapd and direct reclaim was awake longer and for a larger percentage of the overall workload. It's possible there were big differences in the amount of time spent reclaiming slab pages between the different kernels which is plausible considering that the micro tests runs after fsmark and sysbench. Trace Reclaim Statistics: congestion_wait Direct number congest waited 845 1312 104 0 Direct time congest waited 19416ms 26560ms 7544ms 0ms Direct full congest waited 745 1105 72 0 Direct number conditional waited 0 0 1322 2935 Direct time conditional waited 0ms 0ms 12ms 312ms Direct full conditional waited 0 0 0 3 KSwapd number congest waited 39 102 75 63 KSwapd time congest waited 2484ms 6760ms 5756ms 3716ms KSwapd full congest waited 20 48 46 25 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 20 seconds asleep in direct reclaim and only 312ms asleep with the patches. The time kswapd spent congest waited was also reduced by a large factor. MMTests Statistics: duration ser/Sys Time Running Test (seconds) 26.58 28.05 26.9 28.47 Total Elapsed Time (seconds) 30.02 32.86 32.67 30.88 With all patches applies, the completion times are very similar. X86-64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 82.00 ( 0.00%) 84.00 ( 2.00%) 85.00 ( 3.00%) 85.00 ( 3.00%) Pass 2 90.00 ( 0.00%) 87.00 (-3.00%) 88.00 (-2.00%) 89.00 (-1.00%) At Rest 92.00 ( 0.00%) 90.00 (-2.00%) 90.00 (-2.00%) 91.00 (-1.00%) Success figures across the board are broadly similar. traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 1045 944 886 887 Direct reclaim pages scanned 135091 119604 109382 101019 Direct reclaim pages reclaimed 88599 47535 47863 46671 Direct reclaim write file async I/O 494 283 465 280 Direct reclaim write anon async I/O 29357 13710 16656 13462 Direct reclaim write file sync I/O 154 2 2 3 Direct reclaim write anon sync I/O 14594 571 509 561 Wake kswapd requests 7491 933 872 892 Kswapd wakeups 814 778 731 780 Kswapd pages scanned 7290822 15341158 11916436 13703442 Kswapd pages reclaimed 3587336 3142496 3094392 3187151 Kswapd reclaim write file async I/O 91975 32317 28022 29628 Kswapd reclaim write anon async I/O 1992022 789307 829745 849769 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4588.93 2467.16 2495.41 2547.07 Time kswapd awake (seconds) 2497.66 1020.16 1098.06 1176.82 Total pages scanned 7425913 15460762 12025818 13804461 Total pages reclaimed 3675935 3190031 3142255 3233822 %age total pages scanned/reclaimed 49.50% 20.63% 26.13% 23.43% %age total pages scanned/written 28.66% 5.41% 7.28% 6.47% %age file pages scanned/written 1.25% 0.21% 0.24% 0.22% Percentage Time Spent Direct Reclaim 57.33% 42.15% 42.41% 42.99% Percentage Time kswapd Awake 43.56% 27.87% 29.76% 31.25% Scanned/reclaimed ratios again look good with big improvements in efficiency. The Scanned/written ratios also look much improved. With a better scanned/written ration, there is an expectation that IO would be more efficient and indeed, the time spent in direct reclaim is much reduced by the full series and kswapd spends a little less time awake. Overall, indications here are that allocations were happening much faster and this can be seen with a graph of the latency figures as the allocations were taking place http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-hydra-mean.ps FTrace Reclaim Statistics: congestion_wait Direct number congest waited 1333 204 169 4 Direct time congest waited 78896ms 8288ms 7260ms 200ms Direct full congest waited 756 92 69 2 Direct number conditional waited 0 0 26 186 Direct time conditional waited 0ms 0ms 0ms 2504ms Direct full conditional waited 0 0 0 25 KSwapd number congest waited 4 395 227 282 KSwapd time congest waited 384ms 25136ms 10508ms 18380ms KSwapd full congest waited 3 232 98 176 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 KSwapd full conditional waited 318 0 312 9 Overall, the time spent speeping is reduced. kswapd is still hitting congestion_wait() but that is because there are callers remaining where it wasn't clear in advance if they should be changed to wait_iff_congested() or not. Overall the sleep imes are reduced though - from 79ish seconds to about 19. MMTests Statistics: duration User/Sys Time Running Test (seconds) 3415.43 3386.65 3388.39 3377.5 Total Elapsed Time (seconds) 5733.48 3660.33 3689.41 3765.39 With the full series, the time to complete the tests are reduced by 30% PPC64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 17.00 ( 0.00%) 34.00 (17.00%) 38.00 (21.00%) 43.00 (26.00%) Pass 2 25.00 ( 0.00%) 37.00 (12.00%) 42.00 (17.00%) 46.00 (21.00%) At Rest 49.00 ( 0.00%) 43.00 (-6.00%) 45.00 (-4.00%) 51.00 ( 2.00%) Success rates there are *way* up particularly considering that the 16MB huge pages on PPC64 mean that it's always much harder to allocate them. FTrace Reclaim Statistics: vmscan stress-highalloc stress-highalloc stress-highalloc stress-highalloc traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 499 505 564 509 Direct reclaim pages scanned 223478 41898 51818 45605 Direct reclaim pages reclaimed 137730 21148 27161 23455 Direct reclaim write file async I/O 399 136 162 136 Direct reclaim write anon async I/O 46977 2865 4686 3998 Direct reclaim write file sync I/O 29 0 1 3 Direct reclaim write anon sync I/O 31023 159 237 239 Wake kswapd requests 420 351 360 326 Kswapd wakeups 185 294 249 277 Kswapd pages scanned 15703488 16392500 17821724 17598737 Kswapd pages reclaimed 5808466 2908858 3139386 3145435 Kswapd reclaim write file async I/O 159938 18400 18717 13473 Kswapd reclaim write anon async I/O 3467554 228957 322799 234278 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 9665.35 1707.81 2374.32 1871.23 Time kswapd awake (seconds) 9401.21 1367.86 1951.75 1328.88 Total pages scanned 15926966 16434398 17873542 17644342 Total pages reclaimed 5946196 2930006 3166547 3168890 %age total pages scanned/reclaimed 37.33% 17.83% 17.72% 17.96% %age total pages scanned/written 23.27% 1.52% 1.94% 1.43% %age file pages scanned/written 1.01% 0.11% 0.11% 0.08% Percentage Time Spent Direct Reclaim 44.55% 35.10% 41.42% 36.91% Percentage Time kswapd Awake 86.71% 43.58% 52.67% 41.14% While the scanning rates are slightly up, the scanned/reclaimed and scanned/written figures are much improved. The time spent in direct reclaim and with kswapd are massively reduced, mostly by the lowlumpy patches. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 725 303 126 3 Direct time congest waited 45524ms 9180ms 5936ms 300ms Direct full congest waited 487 190 52 3 Direct number conditional waited 0 0 200 301 Direct time conditional waited 0ms 0ms 0ms 1904ms Direct full conditional waited 0 0 0 19 KSwapd number congest waited 0 2 23 4 KSwapd time congest waited 0ms 200ms 420ms 404ms KSwapd full congest waited 0 2 2 4 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Not as dramatic a story here but the time spent asleep is reduced and we can still see what wait_iff_congested is going to sleep when necessary. MMTests Statistics: duration User/Sys Time Running Test (seconds) 12028.09 3157.17 3357.79 3199.16 Total Elapsed Time (seconds) 10842.07 3138.72 3705.54 3229.85 The time to complete this test goes way down. With the full series, we are allocating over twice the number of huge pages in 30% of the time and there is a corresponding impact on the allocation latency graph available at. http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-powyah-mean.ps This patch: Add a trace event for shrink_inactive_list() and updates the sample postprocessing script appropriately. It can be used to determine how many pages were reclaimed and for non-lumpy reclaim where exactly the pages were reclaimed from. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-27 05:21:40 +08:00
)
TRACE_EVENT(mm_vmscan_kswapd_sleep,
TP_PROTO(int nid),
TP_ARGS(nid),
TP_STRUCT__entry(
__field( int, nid )
),
TP_fast_assign(
__entry->nid = nid;
),
TP_printk("nid=%d", __entry->nid)
);
TRACE_EVENT(mm_vmscan_kswapd_wake,
TP_PROTO(int nid, int order),
TP_ARGS(nid, order),
TP_STRUCT__entry(
__field( int, nid )
__field( int, order )
),
TP_fast_assign(
__entry->nid = nid;
__entry->order = order;
),
TP_printk("nid=%d order=%d", __entry->nid, __entry->order)
);
TRACE_EVENT(mm_vmscan_wakeup_kswapd,
TP_PROTO(int nid, int zid, int order),
TP_ARGS(nid, zid, order),
TP_STRUCT__entry(
__field( int, nid )
__field( int, zid )
__field( int, order )
),
TP_fast_assign(
__entry->nid = nid;
__entry->zid = zid;
__entry->order = order;
),
TP_printk("nid=%d zid=%d order=%d",
__entry->nid,
__entry->zid,
__entry->order)
);
DECLARE_EVENT_CLASS(mm_vmscan_direct_reclaim_begin_template,
TP_PROTO(int order, int may_writepage, gfp_t gfp_flags),
TP_ARGS(order, may_writepage, gfp_flags),
TP_STRUCT__entry(
__field( int, order )
__field( int, may_writepage )
__field( gfp_t, gfp_flags )
),
TP_fast_assign(
__entry->order = order;
__entry->may_writepage = may_writepage;
__entry->gfp_flags = gfp_flags;
),
TP_printk("order=%d may_writepage=%d gfp_flags=%s",
__entry->order,
__entry->may_writepage,
show_gfp_flags(__entry->gfp_flags))
);
DEFINE_EVENT(mm_vmscan_direct_reclaim_begin_template, mm_vmscan_direct_reclaim_begin,
TP_PROTO(int order, int may_writepage, gfp_t gfp_flags),
TP_ARGS(order, may_writepage, gfp_flags)
);
DEFINE_EVENT(mm_vmscan_direct_reclaim_begin_template, mm_vmscan_memcg_reclaim_begin,
TP_PROTO(int order, int may_writepage, gfp_t gfp_flags),
TP_ARGS(order, may_writepage, gfp_flags)
);
DEFINE_EVENT(mm_vmscan_direct_reclaim_begin_template, mm_vmscan_memcg_softlimit_reclaim_begin,
TP_PROTO(int order, int may_writepage, gfp_t gfp_flags),
TP_ARGS(order, may_writepage, gfp_flags)
);
DECLARE_EVENT_CLASS(mm_vmscan_direct_reclaim_end_template,
TP_PROTO(unsigned long nr_reclaimed),
TP_ARGS(nr_reclaimed),
TP_STRUCT__entry(
__field( unsigned long, nr_reclaimed )
),
TP_fast_assign(
__entry->nr_reclaimed = nr_reclaimed;
),
TP_printk("nr_reclaimed=%lu", __entry->nr_reclaimed)
);
DEFINE_EVENT(mm_vmscan_direct_reclaim_end_template, mm_vmscan_direct_reclaim_end,
TP_PROTO(unsigned long nr_reclaimed),
TP_ARGS(nr_reclaimed)
);
DEFINE_EVENT(mm_vmscan_direct_reclaim_end_template, mm_vmscan_memcg_reclaim_end,
TP_PROTO(unsigned long nr_reclaimed),
TP_ARGS(nr_reclaimed)
);
DEFINE_EVENT(mm_vmscan_direct_reclaim_end_template, mm_vmscan_memcg_softlimit_reclaim_end,
TP_PROTO(unsigned long nr_reclaimed),
TP_ARGS(nr_reclaimed)
);
TRACE_EVENT(mm_shrink_slab_start,
TP_PROTO(struct shrinker *shr, struct shrink_control *sc,
long nr_objects_to_shrink, unsigned long pgs_scanned,
unsigned long lru_pgs, unsigned long cache_items,
unsigned long long delta, unsigned long total_scan),
TP_ARGS(shr, sc, nr_objects_to_shrink, pgs_scanned, lru_pgs,
cache_items, delta, total_scan),
TP_STRUCT__entry(
__field(struct shrinker *, shr)
__field(void *, shrink)
__field(int, nid)
__field(long, nr_objects_to_shrink)
__field(gfp_t, gfp_flags)
__field(unsigned long, pgs_scanned)
__field(unsigned long, lru_pgs)
__field(unsigned long, cache_items)
__field(unsigned long long, delta)
__field(unsigned long, total_scan)
),
TP_fast_assign(
__entry->shr = shr;
shrinker: Kill old ->shrink API. There are no more users of this API, so kill it dead, dead, dead and quietly bury the corpse in a shallow, unmarked grave in a dark forest deep in the hills... [glommer@openvz.org: added flowers to the grave] Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Glauber Costa <glommer@openvz.org> Reviewed-by: Greg Thelen <gthelen@google.com> Acked-by: Mel Gorman <mgorman@suse.de> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com> Cc: Arve Hjønnevåg <arve@android.com> Cc: Carlos Maiolino <cmaiolino@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Cc: David Rientjes <rientjes@google.com> Cc: Gleb Natapov <gleb@redhat.com> Cc: Greg Thelen <gthelen@google.com> Cc: J. Bruce Fields <bfields@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Jerome Glisse <jglisse@redhat.com> Cc: John Stultz <john.stultz@linaro.org> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Kent Overstreet <koverstreet@google.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Thomas Hellstrom <thellstrom@vmware.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 08:18:16 +08:00
__entry->shrink = shr->scan_objects;
__entry->nid = sc->nid;
__entry->nr_objects_to_shrink = nr_objects_to_shrink;
__entry->gfp_flags = sc->gfp_mask;
__entry->pgs_scanned = pgs_scanned;
__entry->lru_pgs = lru_pgs;
__entry->cache_items = cache_items;
__entry->delta = delta;
__entry->total_scan = total_scan;
),
TP_printk("%pF %p: nid: %d objects to shrink %ld gfp_flags %s pgs_scanned %ld lru_pgs %ld cache items %ld delta %lld total_scan %ld",
__entry->shrink,
__entry->shr,
__entry->nid,
__entry->nr_objects_to_shrink,
show_gfp_flags(__entry->gfp_flags),
__entry->pgs_scanned,
__entry->lru_pgs,
__entry->cache_items,
__entry->delta,
__entry->total_scan)
);
TRACE_EVENT(mm_shrink_slab_end,
TP_PROTO(struct shrinker *shr, int nid, int shrinker_retval,
mm: shrinker trace points: fix negatives I was looking at a trace of the slab shrinkers (attachment in this comment): https://bugs.freedesktop.org/show_bug.cgi?id=72742#c67 and noticed that "total_scan" can go negative in some cases. We used to dump out the "total_scan" variable directly, but some of the shrinker modifications along the way changed that. This patch just dumps it out directly, again. It doesn't make any sense to derive it from new_nr and nr any more since there are now other shrinkers that can be running in parallel and mucking with those values. Here's an example of the negative numbers in the output: > kswapd0-840 [000] 160.869398: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 10 new scan count 39 total_scan 29 last shrinker return val 256 > kswapd0-840 [000] 160.869618: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 39 new scan count 102 total_scan 63 last shrinker return val 256 > kswapd0-840 [000] 160.870031: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 102 new scan count 47 total_scan -55 last shrinker return val 768 > kswapd0-840 [000] 160.870464: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 47 new scan count 45 total_scan -2 last shrinker return val 768 > kswapd0-840 [000] 163.384144: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 45 new scan count 56 total_scan 11 last shrinker return val 0 > kswapd0-840 [000] 163.384297: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 56 new scan count 15 total_scan -41 last shrinker return val 256 > kswapd0-840 [000] 163.384414: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 15 new scan count 117 total_scan 102 last shrinker return val 0 > kswapd0-840 [000] 163.384657: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 117 new scan count 36 total_scan -81 last shrinker return val 512 > kswapd0-840 [000] 163.384880: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 36 new scan count 111 total_scan 75 last shrinker return val 256 > kswapd0-840 [000] 163.385256: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 111 new scan count 34 total_scan -77 last shrinker return val 768 > kswapd0-840 [000] 163.385598: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 34 new scan count 122 total_scan 88 last shrinker return val 512 Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Dave Chinner <david@fromorbit.com> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-05 07:08:06 +08:00
long unused_scan_cnt, long new_scan_cnt, long total_scan),
TP_ARGS(shr, nid, shrinker_retval, unused_scan_cnt, new_scan_cnt,
total_scan),
TP_STRUCT__entry(
__field(struct shrinker *, shr)
__field(int, nid)
__field(void *, shrink)
__field(long, unused_scan)
__field(long, new_scan)
__field(int, retval)
__field(long, total_scan)
),
TP_fast_assign(
__entry->shr = shr;
__entry->nid = nid;
shrinker: Kill old ->shrink API. There are no more users of this API, so kill it dead, dead, dead and quietly bury the corpse in a shallow, unmarked grave in a dark forest deep in the hills... [glommer@openvz.org: added flowers to the grave] Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Glauber Costa <glommer@openvz.org> Reviewed-by: Greg Thelen <gthelen@google.com> Acked-by: Mel Gorman <mgorman@suse.de> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com> Cc: Arve Hjønnevåg <arve@android.com> Cc: Carlos Maiolino <cmaiolino@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Cc: David Rientjes <rientjes@google.com> Cc: Gleb Natapov <gleb@redhat.com> Cc: Greg Thelen <gthelen@google.com> Cc: J. Bruce Fields <bfields@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Jerome Glisse <jglisse@redhat.com> Cc: John Stultz <john.stultz@linaro.org> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Kent Overstreet <koverstreet@google.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Mel Gorman <mgorman@suse.de> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Thomas Hellstrom <thellstrom@vmware.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 08:18:16 +08:00
__entry->shrink = shr->scan_objects;
__entry->unused_scan = unused_scan_cnt;
__entry->new_scan = new_scan_cnt;
__entry->retval = shrinker_retval;
mm: shrinker trace points: fix negatives I was looking at a trace of the slab shrinkers (attachment in this comment): https://bugs.freedesktop.org/show_bug.cgi?id=72742#c67 and noticed that "total_scan" can go negative in some cases. We used to dump out the "total_scan" variable directly, but some of the shrinker modifications along the way changed that. This patch just dumps it out directly, again. It doesn't make any sense to derive it from new_nr and nr any more since there are now other shrinkers that can be running in parallel and mucking with those values. Here's an example of the negative numbers in the output: > kswapd0-840 [000] 160.869398: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 10 new scan count 39 total_scan 29 last shrinker return val 256 > kswapd0-840 [000] 160.869618: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 39 new scan count 102 total_scan 63 last shrinker return val 256 > kswapd0-840 [000] 160.870031: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 102 new scan count 47 total_scan -55 last shrinker return val 768 > kswapd0-840 [000] 160.870464: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 47 new scan count 45 total_scan -2 last shrinker return val 768 > kswapd0-840 [000] 163.384144: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 45 new scan count 56 total_scan 11 last shrinker return val 0 > kswapd0-840 [000] 163.384297: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 56 new scan count 15 total_scan -41 last shrinker return val 256 > kswapd0-840 [000] 163.384414: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 15 new scan count 117 total_scan 102 last shrinker return val 0 > kswapd0-840 [000] 163.384657: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 117 new scan count 36 total_scan -81 last shrinker return val 512 > kswapd0-840 [000] 163.384880: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 36 new scan count 111 total_scan 75 last shrinker return val 256 > kswapd0-840 [000] 163.385256: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 111 new scan count 34 total_scan -77 last shrinker return val 768 > kswapd0-840 [000] 163.385598: mm_shrink_slab_end: i915_gem_inactive_scan+0x0 0xffff8800037cbc68: unused scan count 34 new scan count 122 total_scan 88 last shrinker return val 512 Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Dave Chinner <david@fromorbit.com> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-05 07:08:06 +08:00
__entry->total_scan = total_scan;
),
TP_printk("%pF %p: nid: %d unused scan count %ld new scan count %ld total_scan %ld last shrinker return val %d",
__entry->shrink,
__entry->shr,
__entry->nid,
__entry->unused_scan,
__entry->new_scan,
__entry->total_scan,
__entry->retval)
);
DECLARE_EVENT_CLASS(mm_vmscan_lru_isolate_template,
TP_PROTO(int order,
unsigned long nr_requested,
unsigned long nr_scanned,
unsigned long nr_taken,
isolate_mode_t isolate_mode,
int file),
mm: vmscan: remove lumpy reclaim This series removes lumpy reclaim and some stalling logic that was unintentionally being used by memory compaction. The end result is that stalling on dirty pages during page reclaim now depends on wait_iff_congested(). Four kernels were compared 3.3.0 vanilla 3.4.0-rc2 vanilla 3.4.0-rc2 lumpyremove-v2 is patch one from this series 3.4.0-rc2 nosync-v2r3 is the full series Removing lumpy reclaim saves almost 900 bytes of text whereas the full series removes 1200 bytes. text data bss dec hex filename 6740375 1927944 2260992 10929311 a6c49f vmlinux-3.4.0-rc2-vanilla 6739479 1927944 2260992 10928415 a6c11f vmlinux-3.4.0-rc2-lumpyremove-v2 6739159 1927944 2260992 10928095 a6bfdf vmlinux-3.4.0-rc2-nosync-v2 There are behaviour changes in the series and so tests were run with monitoring of ftrace events. This disrupts results so the performance results are distorted but the new behaviour should be clearer. fs-mark running in a threaded configuration showed little of interest as it did not push reclaim aggressively FS-Mark Multi Threaded 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Files/s min 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s mean 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s stddev 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Files/s max 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Overhead min 508667.00 ( 0.00%) 521350.00 (-2.49%) 544292.00 (-7.00%) 547168.00 (-7.57%) Overhead mean 551185.00 ( 0.00%) 652690.73 (-18.42%) 991208.40 (-79.83%) 570130.53 (-3.44%) Overhead stddev 18200.69 ( 0.00%) 331958.29 (-1723.88%) 1579579.43 (-8578.68%) 9576.81 (47.38%) Overhead max 576775.00 ( 0.00%) 1846634.00 (-220.17%) 6901055.00 (-1096.49%) 585675.00 (-1.54%) MMTests Statistics: duration Sys Time Running Test (seconds) 309.90 300.95 307.33 298.95 User+Sys Time Running Test (seconds) 319.32 309.67 315.69 307.51 Total Elapsed Time (seconds) 1187.85 1193.09 1191.98 1193.73 MMTests Statistics: vmstat Page Ins 80532 82212 81420 79480 Page Outs 111434984 111456240 111437376 111582628 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 44881 27889 27453 34843 Kswapd pages scanned 25841428 25860774 25861233 25843212 Kswapd pages reclaimed 25841393 25860741 25861199 25843179 Direct pages reclaimed 44881 27889 27453 34843 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 21754.791 21675.460 21696.029 21649.127 Direct efficiency 100% 100% 100% 100% Direct velocity 37.783 23.375 23.031 29.188 Percentage direct scans 0% 0% 0% 0% ftrace showed that there was no stalling on writeback or pages submitted for IO from reclaim context. postmark was similar and while it was more interesting, it also did not push reclaim heavily. POSTMARK 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Transactions per second: 16.00 ( 0.00%) 20.00 (25.00%) 18.00 (12.50%) 17.00 ( 6.25%) Data megabytes read per second: 18.80 ( 0.00%) 24.27 (29.10%) 22.26 (18.40%) 20.54 ( 9.26%) Data megabytes written per second: 35.83 ( 0.00%) 46.25 (29.08%) 42.42 (18.39%) 39.14 ( 9.24%) Files created alone per second: 28.00 ( 0.00%) 38.00 (35.71%) 34.00 (21.43%) 30.00 ( 7.14%) Files create/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) Files deleted alone per second: 556.00 ( 0.00%) 1224.00 (120.14%) 3062.00 (450.72%) 6124.00 (1001.44%) Files delete/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 113.34 107.99 109.73 108.72 User+Sys Time Running Test (seconds) 145.51 139.81 143.32 143.55 Total Elapsed Time (seconds) 1159.16 899.23 980.17 1062.27 MMTests Statistics: vmstat Page Ins 13710192 13729032 13727944 13760136 Page Outs 43071140 42987228 42733684 42931624 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 0 0 0 0 Kswapd pages scanned 9941613 9937443 9939085 9929154 Kswapd pages reclaimed 9940926 9936751 9938397 9928465 Direct pages reclaimed 0 0 0 0 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 8576.567 11051.058 10140.164 9347.109 Direct efficiency 100% 100% 100% 100% Direct velocity 0.000 0.000 0.000 0.000 It looks like here that the full series regresses performance but as ftrace showed no usage of wait_iff_congested() or sync reclaim I am assuming it's a disruption due to monitoring. Other data such as memory usage, page IO, swap IO all looked similar. Running a benchmark with a plain DD showed nothing very interesting. The full series stalled in wait_iff_congested() slightly less but stall times on vanilla kernels were marginal. Running a benchmark that hammered on file-backed mappings showed stalls due to congestion but not in sync writebacks MICRO 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 MMTests Statistics: duration Sys Time Running Test (seconds) 308.13 294.50 298.75 299.53 User+Sys Time Running Test (seconds) 330.45 316.28 318.93 320.79 Total Elapsed Time (seconds) 1814.90 1833.88 1821.14 1832.91 MMTests Statistics: vmstat Page Ins 108712 120708 97224 110344 Page Outs 155514576 156017404 155813676 156193256 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 2599253 1550480 2512822 2414760 Kswapd pages scanned 69742364 71150694 68839041 69692533 Kswapd pages reclaimed 34824488 34773341 34796602 34799396 Direct pages reclaimed 53693 94750 61792 75205 Kswapd efficiency 49% 48% 50% 49% Kswapd velocity 38427.662 38797.901 37799.972 38022.889 Direct efficiency 2% 6% 2% 3% Direct velocity 1432.174 845.464 1379.807 1317.446 Percentage direct scans 3% 2% 3% 3% Page writes by reclaim 0 0 0 0 Page writes file 0 0 0 0 Page writes anon 0 0 0 0 Page reclaim immediate 0 0 0 1218 Page rescued immediate 0 0 0 0 Slabs scanned 15360 16384 13312 16384 Direct inode steals 0 0 0 0 Kswapd inode steals 4340 4327 1630 4323 FTrace Reclaim Statistics: congestion_wait Direct number congest waited 0 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 900 870 754 789 Direct time conditional waited 0ms 0ms 0ms 20ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 2106 2308 2116 1915 KSwapd time congest waited 139924ms 157832ms 125652ms 132516ms KSwapd full congest waited 1346 1530 1202 1278 KSwapd number conditional waited 12922 16320 10943 14670 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Reclaim statistics are not radically changed. The stall times in kswapd are massive but it is clear that it is due to calls to congestion_wait() and that is almost certainly the call in balance_pgdat(). Otherwise stalls due to dirty pages are non-existant. I ran a benchmark that stressed high-order allocation. This is very artifical load but was used in the past to evaluate lumpy reclaim and compaction. Generally I look at allocation success rates and latency figures. STRESS-HIGHALLOC 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Pass 1 81.00 ( 0.00%) 28.00 (-53.00%) 24.00 (-57.00%) 28.00 (-53.00%) Pass 2 82.00 ( 0.00%) 39.00 (-43.00%) 38.00 (-44.00%) 43.00 (-39.00%) while Rested 88.00 ( 0.00%) 87.00 (-1.00%) 88.00 ( 0.00%) 88.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 740.93 681.42 685.14 684.87 User+Sys Time Running Test (seconds) 2922.65 3269.52 3281.35 3279.44 Total Elapsed Time (seconds) 1161.73 1152.49 1159.55 1161.44 MMTests Statistics: vmstat Page Ins 4486020 2807256 2855944 2876244 Page Outs 7261600 7973688 7975320 7986120 Swap Ins 31694 0 0 0 Swap Outs 98179 0 0 0 Direct pages scanned 53494 57731 34406 113015 Kswapd pages scanned 6271173 1287481 1278174 1219095 Kswapd pages reclaimed 2029240 1281025 1260708 1201583 Direct pages reclaimed 1468 14564 16649 92456 Kswapd efficiency 32% 99% 98% 98% Kswapd velocity 5398.133 1117.130 1102.302 1049.641 Direct efficiency 2% 25% 48% 81% Direct velocity 46.047 50.092 29.672 97.306 Percentage direct scans 0% 4% 2% 8% Page writes by reclaim 1616049 0 0 0 Page writes file 1517870 0 0 0 Page writes anon 98179 0 0 0 Page reclaim immediate 103778 27339 9796 17831 Page rescued immediate 0 0 0 0 Slabs scanned 1096704 986112 980992 998400 Direct inode steals 223 215040 216736 247881 Kswapd inode steals 175331 61548 68444 63066 Kswapd skipped wait 21991 0 1 0 THP fault alloc 1 135 125 134 THP collapse alloc 393 311 228 236 THP splits 25 13 7 8 THP fault fallback 0 0 0 0 THP collapse fail 3 5 7 7 Compaction stalls 865 1270 1422 1518 Compaction success 370 401 353 383 Compaction failures 495 869 1069 1135 Compaction pages moved 870155 3828868 4036106 4423626 Compaction move failure 26429 23865 29742 27514 Success rates are completely hosed for 3.4-rc2 which is almost certainly due to commit fe2c2a106663 ("vmscan: reclaim at order 0 when compaction is enabled"). I expected this would happen for kswapd and impair allocation success rates (https://lkml.org/lkml/2012/1/25/166) but I did not anticipate this much a difference: 80% less scanning, 37% less reclaim by kswapd In comparison, reclaim/compaction is not aggressive and gives up easily which is the intended behaviour. hugetlbfs uses __GFP_REPEAT and would be much more aggressive about reclaim/compaction than THP allocations are. The stress test above is allocating like neither THP or hugetlbfs but is much closer to THP. Mainline is now impaired in terms of high order allocation under heavy load although I do not know to what degree as I did not test with __GFP_REPEAT. Keep this in mind for bugs related to hugepage pool resizing, THP allocation and high order atomic allocation failures from network devices. In terms of congestion throttling, I see the following for this test FTrace Reclaim Statistics: congestion_wait Direct number congest waited 3 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 957 512 1081 1075 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 36 4 3 5 KSwapd time congest waited 3148ms 400ms 300ms 500ms KSwapd full congest waited 30 4 3 5 KSwapd number conditional waited 88514 197 332 542 KSwapd time conditional waited 4980ms 0ms 0ms 0ms KSwapd full conditional waited 49 0 0 0 The "conditional waited" times are the most interesting as this is directly impacted by the number of dirty pages encountered during scan. As lumpy reclaim is no longer scanning contiguous ranges, it is finding fewer dirty pages. This brings wait times from about 5 seconds to 0. kswapd itself is still calling congestion_wait() so it'll still stall but it's a lot less. In terms of the type of IO we were doing, I see this FTrace Reclaim Statistics: mm_vmscan_writepage Direct writes anon sync 0 0 0 0 Direct writes anon async 0 0 0 0 Direct writes file sync 0 0 0 0 Direct writes file async 0 0 0 0 Direct writes mixed sync 0 0 0 0 Direct writes mixed async 0 0 0 0 KSwapd writes anon sync 0 0 0 0 KSwapd writes anon async 91682 0 0 0 KSwapd writes file sync 0 0 0 0 KSwapd writes file async 822629 0 0 0 KSwapd writes mixed sync 0 0 0 0 KSwapd writes mixed async 0 0 0 0 In 3.2, kswapd was doing a bunch of async writes of pages but reclaim/compaction was never reaching a point where it was doing sync IO. This does not guarantee that reclaim/compaction was not calling wait_on_page_writeback() but I would consider it unlikely. It indicates that merging patches 2 and 3 to stop reclaim/compaction calling wait_on_page_writeback() should be safe. This patch: Lumpy reclaim had a purpose but in the mind of some, it was to kick the system so hard it trashed. For others the purpose was to complicate vmscan.c. Over time it was giving softer shoes and a nicer attitude but memory compaction needs to step up and replace it so this patch sends lumpy reclaim to the farm. The tracepoint format changes for isolating LRU pages with this patch applied. Furthermore reclaim/compaction can no longer queue dirty pages in pageout() if the underlying BDI is congested. Lumpy reclaim used this logic and reclaim/compaction was using it in error. Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Cc: Hugh Dickins <hughd@google.com> Cc: Ying Han <yinghan@google.com> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-30 06:06:19 +08:00
TP_ARGS(order, nr_requested, nr_scanned, nr_taken, isolate_mode, file),
TP_STRUCT__entry(
__field(int, order)
__field(unsigned long, nr_requested)
__field(unsigned long, nr_scanned)
__field(unsigned long, nr_taken)
__field(isolate_mode_t, isolate_mode)
__field(int, file)
),
TP_fast_assign(
__entry->order = order;
__entry->nr_requested = nr_requested;
__entry->nr_scanned = nr_scanned;
__entry->nr_taken = nr_taken;
__entry->isolate_mode = isolate_mode;
__entry->file = file;
),
mm: vmscan: remove lumpy reclaim This series removes lumpy reclaim and some stalling logic that was unintentionally being used by memory compaction. The end result is that stalling on dirty pages during page reclaim now depends on wait_iff_congested(). Four kernels were compared 3.3.0 vanilla 3.4.0-rc2 vanilla 3.4.0-rc2 lumpyremove-v2 is patch one from this series 3.4.0-rc2 nosync-v2r3 is the full series Removing lumpy reclaim saves almost 900 bytes of text whereas the full series removes 1200 bytes. text data bss dec hex filename 6740375 1927944 2260992 10929311 a6c49f vmlinux-3.4.0-rc2-vanilla 6739479 1927944 2260992 10928415 a6c11f vmlinux-3.4.0-rc2-lumpyremove-v2 6739159 1927944 2260992 10928095 a6bfdf vmlinux-3.4.0-rc2-nosync-v2 There are behaviour changes in the series and so tests were run with monitoring of ftrace events. This disrupts results so the performance results are distorted but the new behaviour should be clearer. fs-mark running in a threaded configuration showed little of interest as it did not push reclaim aggressively FS-Mark Multi Threaded 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Files/s min 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s mean 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s stddev 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Files/s max 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Overhead min 508667.00 ( 0.00%) 521350.00 (-2.49%) 544292.00 (-7.00%) 547168.00 (-7.57%) Overhead mean 551185.00 ( 0.00%) 652690.73 (-18.42%) 991208.40 (-79.83%) 570130.53 (-3.44%) Overhead stddev 18200.69 ( 0.00%) 331958.29 (-1723.88%) 1579579.43 (-8578.68%) 9576.81 (47.38%) Overhead max 576775.00 ( 0.00%) 1846634.00 (-220.17%) 6901055.00 (-1096.49%) 585675.00 (-1.54%) MMTests Statistics: duration Sys Time Running Test (seconds) 309.90 300.95 307.33 298.95 User+Sys Time Running Test (seconds) 319.32 309.67 315.69 307.51 Total Elapsed Time (seconds) 1187.85 1193.09 1191.98 1193.73 MMTests Statistics: vmstat Page Ins 80532 82212 81420 79480 Page Outs 111434984 111456240 111437376 111582628 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 44881 27889 27453 34843 Kswapd pages scanned 25841428 25860774 25861233 25843212 Kswapd pages reclaimed 25841393 25860741 25861199 25843179 Direct pages reclaimed 44881 27889 27453 34843 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 21754.791 21675.460 21696.029 21649.127 Direct efficiency 100% 100% 100% 100% Direct velocity 37.783 23.375 23.031 29.188 Percentage direct scans 0% 0% 0% 0% ftrace showed that there was no stalling on writeback or pages submitted for IO from reclaim context. postmark was similar and while it was more interesting, it also did not push reclaim heavily. POSTMARK 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Transactions per second: 16.00 ( 0.00%) 20.00 (25.00%) 18.00 (12.50%) 17.00 ( 6.25%) Data megabytes read per second: 18.80 ( 0.00%) 24.27 (29.10%) 22.26 (18.40%) 20.54 ( 9.26%) Data megabytes written per second: 35.83 ( 0.00%) 46.25 (29.08%) 42.42 (18.39%) 39.14 ( 9.24%) Files created alone per second: 28.00 ( 0.00%) 38.00 (35.71%) 34.00 (21.43%) 30.00 ( 7.14%) Files create/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) Files deleted alone per second: 556.00 ( 0.00%) 1224.00 (120.14%) 3062.00 (450.72%) 6124.00 (1001.44%) Files delete/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 113.34 107.99 109.73 108.72 User+Sys Time Running Test (seconds) 145.51 139.81 143.32 143.55 Total Elapsed Time (seconds) 1159.16 899.23 980.17 1062.27 MMTests Statistics: vmstat Page Ins 13710192 13729032 13727944 13760136 Page Outs 43071140 42987228 42733684 42931624 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 0 0 0 0 Kswapd pages scanned 9941613 9937443 9939085 9929154 Kswapd pages reclaimed 9940926 9936751 9938397 9928465 Direct pages reclaimed 0 0 0 0 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 8576.567 11051.058 10140.164 9347.109 Direct efficiency 100% 100% 100% 100% Direct velocity 0.000 0.000 0.000 0.000 It looks like here that the full series regresses performance but as ftrace showed no usage of wait_iff_congested() or sync reclaim I am assuming it's a disruption due to monitoring. Other data such as memory usage, page IO, swap IO all looked similar. Running a benchmark with a plain DD showed nothing very interesting. The full series stalled in wait_iff_congested() slightly less but stall times on vanilla kernels were marginal. Running a benchmark that hammered on file-backed mappings showed stalls due to congestion but not in sync writebacks MICRO 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 MMTests Statistics: duration Sys Time Running Test (seconds) 308.13 294.50 298.75 299.53 User+Sys Time Running Test (seconds) 330.45 316.28 318.93 320.79 Total Elapsed Time (seconds) 1814.90 1833.88 1821.14 1832.91 MMTests Statistics: vmstat Page Ins 108712 120708 97224 110344 Page Outs 155514576 156017404 155813676 156193256 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 2599253 1550480 2512822 2414760 Kswapd pages scanned 69742364 71150694 68839041 69692533 Kswapd pages reclaimed 34824488 34773341 34796602 34799396 Direct pages reclaimed 53693 94750 61792 75205 Kswapd efficiency 49% 48% 50% 49% Kswapd velocity 38427.662 38797.901 37799.972 38022.889 Direct efficiency 2% 6% 2% 3% Direct velocity 1432.174 845.464 1379.807 1317.446 Percentage direct scans 3% 2% 3% 3% Page writes by reclaim 0 0 0 0 Page writes file 0 0 0 0 Page writes anon 0 0 0 0 Page reclaim immediate 0 0 0 1218 Page rescued immediate 0 0 0 0 Slabs scanned 15360 16384 13312 16384 Direct inode steals 0 0 0 0 Kswapd inode steals 4340 4327 1630 4323 FTrace Reclaim Statistics: congestion_wait Direct number congest waited 0 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 900 870 754 789 Direct time conditional waited 0ms 0ms 0ms 20ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 2106 2308 2116 1915 KSwapd time congest waited 139924ms 157832ms 125652ms 132516ms KSwapd full congest waited 1346 1530 1202 1278 KSwapd number conditional waited 12922 16320 10943 14670 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Reclaim statistics are not radically changed. The stall times in kswapd are massive but it is clear that it is due to calls to congestion_wait() and that is almost certainly the call in balance_pgdat(). Otherwise stalls due to dirty pages are non-existant. I ran a benchmark that stressed high-order allocation. This is very artifical load but was used in the past to evaluate lumpy reclaim and compaction. Generally I look at allocation success rates and latency figures. STRESS-HIGHALLOC 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Pass 1 81.00 ( 0.00%) 28.00 (-53.00%) 24.00 (-57.00%) 28.00 (-53.00%) Pass 2 82.00 ( 0.00%) 39.00 (-43.00%) 38.00 (-44.00%) 43.00 (-39.00%) while Rested 88.00 ( 0.00%) 87.00 (-1.00%) 88.00 ( 0.00%) 88.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 740.93 681.42 685.14 684.87 User+Sys Time Running Test (seconds) 2922.65 3269.52 3281.35 3279.44 Total Elapsed Time (seconds) 1161.73 1152.49 1159.55 1161.44 MMTests Statistics: vmstat Page Ins 4486020 2807256 2855944 2876244 Page Outs 7261600 7973688 7975320 7986120 Swap Ins 31694 0 0 0 Swap Outs 98179 0 0 0 Direct pages scanned 53494 57731 34406 113015 Kswapd pages scanned 6271173 1287481 1278174 1219095 Kswapd pages reclaimed 2029240 1281025 1260708 1201583 Direct pages reclaimed 1468 14564 16649 92456 Kswapd efficiency 32% 99% 98% 98% Kswapd velocity 5398.133 1117.130 1102.302 1049.641 Direct efficiency 2% 25% 48% 81% Direct velocity 46.047 50.092 29.672 97.306 Percentage direct scans 0% 4% 2% 8% Page writes by reclaim 1616049 0 0 0 Page writes file 1517870 0 0 0 Page writes anon 98179 0 0 0 Page reclaim immediate 103778 27339 9796 17831 Page rescued immediate 0 0 0 0 Slabs scanned 1096704 986112 980992 998400 Direct inode steals 223 215040 216736 247881 Kswapd inode steals 175331 61548 68444 63066 Kswapd skipped wait 21991 0 1 0 THP fault alloc 1 135 125 134 THP collapse alloc 393 311 228 236 THP splits 25 13 7 8 THP fault fallback 0 0 0 0 THP collapse fail 3 5 7 7 Compaction stalls 865 1270 1422 1518 Compaction success 370 401 353 383 Compaction failures 495 869 1069 1135 Compaction pages moved 870155 3828868 4036106 4423626 Compaction move failure 26429 23865 29742 27514 Success rates are completely hosed for 3.4-rc2 which is almost certainly due to commit fe2c2a106663 ("vmscan: reclaim at order 0 when compaction is enabled"). I expected this would happen for kswapd and impair allocation success rates (https://lkml.org/lkml/2012/1/25/166) but I did not anticipate this much a difference: 80% less scanning, 37% less reclaim by kswapd In comparison, reclaim/compaction is not aggressive and gives up easily which is the intended behaviour. hugetlbfs uses __GFP_REPEAT and would be much more aggressive about reclaim/compaction than THP allocations are. The stress test above is allocating like neither THP or hugetlbfs but is much closer to THP. Mainline is now impaired in terms of high order allocation under heavy load although I do not know to what degree as I did not test with __GFP_REPEAT. Keep this in mind for bugs related to hugepage pool resizing, THP allocation and high order atomic allocation failures from network devices. In terms of congestion throttling, I see the following for this test FTrace Reclaim Statistics: congestion_wait Direct number congest waited 3 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 957 512 1081 1075 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 36 4 3 5 KSwapd time congest waited 3148ms 400ms 300ms 500ms KSwapd full congest waited 30 4 3 5 KSwapd number conditional waited 88514 197 332 542 KSwapd time conditional waited 4980ms 0ms 0ms 0ms KSwapd full conditional waited 49 0 0 0 The "conditional waited" times are the most interesting as this is directly impacted by the number of dirty pages encountered during scan. As lumpy reclaim is no longer scanning contiguous ranges, it is finding fewer dirty pages. This brings wait times from about 5 seconds to 0. kswapd itself is still calling congestion_wait() so it'll still stall but it's a lot less. In terms of the type of IO we were doing, I see this FTrace Reclaim Statistics: mm_vmscan_writepage Direct writes anon sync 0 0 0 0 Direct writes anon async 0 0 0 0 Direct writes file sync 0 0 0 0 Direct writes file async 0 0 0 0 Direct writes mixed sync 0 0 0 0 Direct writes mixed async 0 0 0 0 KSwapd writes anon sync 0 0 0 0 KSwapd writes anon async 91682 0 0 0 KSwapd writes file sync 0 0 0 0 KSwapd writes file async 822629 0 0 0 KSwapd writes mixed sync 0 0 0 0 KSwapd writes mixed async 0 0 0 0 In 3.2, kswapd was doing a bunch of async writes of pages but reclaim/compaction was never reaching a point where it was doing sync IO. This does not guarantee that reclaim/compaction was not calling wait_on_page_writeback() but I would consider it unlikely. It indicates that merging patches 2 and 3 to stop reclaim/compaction calling wait_on_page_writeback() should be safe. This patch: Lumpy reclaim had a purpose but in the mind of some, it was to kick the system so hard it trashed. For others the purpose was to complicate vmscan.c. Over time it was giving softer shoes and a nicer attitude but memory compaction needs to step up and replace it so this patch sends lumpy reclaim to the farm. The tracepoint format changes for isolating LRU pages with this patch applied. Furthermore reclaim/compaction can no longer queue dirty pages in pageout() if the underlying BDI is congested. Lumpy reclaim used this logic and reclaim/compaction was using it in error. Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Cc: Hugh Dickins <hughd@google.com> Cc: Ying Han <yinghan@google.com> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-30 06:06:19 +08:00
TP_printk("isolate_mode=%d order=%d nr_requested=%lu nr_scanned=%lu nr_taken=%lu file=%d",
__entry->isolate_mode,
__entry->order,
__entry->nr_requested,
__entry->nr_scanned,
__entry->nr_taken,
__entry->file)
);
DEFINE_EVENT(mm_vmscan_lru_isolate_template, mm_vmscan_lru_isolate,
TP_PROTO(int order,
unsigned long nr_requested,
unsigned long nr_scanned,
unsigned long nr_taken,
isolate_mode_t isolate_mode,
int file),
mm: vmscan: remove lumpy reclaim This series removes lumpy reclaim and some stalling logic that was unintentionally being used by memory compaction. The end result is that stalling on dirty pages during page reclaim now depends on wait_iff_congested(). Four kernels were compared 3.3.0 vanilla 3.4.0-rc2 vanilla 3.4.0-rc2 lumpyremove-v2 is patch one from this series 3.4.0-rc2 nosync-v2r3 is the full series Removing lumpy reclaim saves almost 900 bytes of text whereas the full series removes 1200 bytes. text data bss dec hex filename 6740375 1927944 2260992 10929311 a6c49f vmlinux-3.4.0-rc2-vanilla 6739479 1927944 2260992 10928415 a6c11f vmlinux-3.4.0-rc2-lumpyremove-v2 6739159 1927944 2260992 10928095 a6bfdf vmlinux-3.4.0-rc2-nosync-v2 There are behaviour changes in the series and so tests were run with monitoring of ftrace events. This disrupts results so the performance results are distorted but the new behaviour should be clearer. fs-mark running in a threaded configuration showed little of interest as it did not push reclaim aggressively FS-Mark Multi Threaded 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Files/s min 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s mean 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s stddev 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Files/s max 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Overhead min 508667.00 ( 0.00%) 521350.00 (-2.49%) 544292.00 (-7.00%) 547168.00 (-7.57%) Overhead mean 551185.00 ( 0.00%) 652690.73 (-18.42%) 991208.40 (-79.83%) 570130.53 (-3.44%) Overhead stddev 18200.69 ( 0.00%) 331958.29 (-1723.88%) 1579579.43 (-8578.68%) 9576.81 (47.38%) Overhead max 576775.00 ( 0.00%) 1846634.00 (-220.17%) 6901055.00 (-1096.49%) 585675.00 (-1.54%) MMTests Statistics: duration Sys Time Running Test (seconds) 309.90 300.95 307.33 298.95 User+Sys Time Running Test (seconds) 319.32 309.67 315.69 307.51 Total Elapsed Time (seconds) 1187.85 1193.09 1191.98 1193.73 MMTests Statistics: vmstat Page Ins 80532 82212 81420 79480 Page Outs 111434984 111456240 111437376 111582628 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 44881 27889 27453 34843 Kswapd pages scanned 25841428 25860774 25861233 25843212 Kswapd pages reclaimed 25841393 25860741 25861199 25843179 Direct pages reclaimed 44881 27889 27453 34843 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 21754.791 21675.460 21696.029 21649.127 Direct efficiency 100% 100% 100% 100% Direct velocity 37.783 23.375 23.031 29.188 Percentage direct scans 0% 0% 0% 0% ftrace showed that there was no stalling on writeback or pages submitted for IO from reclaim context. postmark was similar and while it was more interesting, it also did not push reclaim heavily. POSTMARK 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Transactions per second: 16.00 ( 0.00%) 20.00 (25.00%) 18.00 (12.50%) 17.00 ( 6.25%) Data megabytes read per second: 18.80 ( 0.00%) 24.27 (29.10%) 22.26 (18.40%) 20.54 ( 9.26%) Data megabytes written per second: 35.83 ( 0.00%) 46.25 (29.08%) 42.42 (18.39%) 39.14 ( 9.24%) Files created alone per second: 28.00 ( 0.00%) 38.00 (35.71%) 34.00 (21.43%) 30.00 ( 7.14%) Files create/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) Files deleted alone per second: 556.00 ( 0.00%) 1224.00 (120.14%) 3062.00 (450.72%) 6124.00 (1001.44%) Files delete/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 113.34 107.99 109.73 108.72 User+Sys Time Running Test (seconds) 145.51 139.81 143.32 143.55 Total Elapsed Time (seconds) 1159.16 899.23 980.17 1062.27 MMTests Statistics: vmstat Page Ins 13710192 13729032 13727944 13760136 Page Outs 43071140 42987228 42733684 42931624 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 0 0 0 0 Kswapd pages scanned 9941613 9937443 9939085 9929154 Kswapd pages reclaimed 9940926 9936751 9938397 9928465 Direct pages reclaimed 0 0 0 0 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 8576.567 11051.058 10140.164 9347.109 Direct efficiency 100% 100% 100% 100% Direct velocity 0.000 0.000 0.000 0.000 It looks like here that the full series regresses performance but as ftrace showed no usage of wait_iff_congested() or sync reclaim I am assuming it's a disruption due to monitoring. Other data such as memory usage, page IO, swap IO all looked similar. Running a benchmark with a plain DD showed nothing very interesting. The full series stalled in wait_iff_congested() slightly less but stall times on vanilla kernels were marginal. Running a benchmark that hammered on file-backed mappings showed stalls due to congestion but not in sync writebacks MICRO 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 MMTests Statistics: duration Sys Time Running Test (seconds) 308.13 294.50 298.75 299.53 User+Sys Time Running Test (seconds) 330.45 316.28 318.93 320.79 Total Elapsed Time (seconds) 1814.90 1833.88 1821.14 1832.91 MMTests Statistics: vmstat Page Ins 108712 120708 97224 110344 Page Outs 155514576 156017404 155813676 156193256 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 2599253 1550480 2512822 2414760 Kswapd pages scanned 69742364 71150694 68839041 69692533 Kswapd pages reclaimed 34824488 34773341 34796602 34799396 Direct pages reclaimed 53693 94750 61792 75205 Kswapd efficiency 49% 48% 50% 49% Kswapd velocity 38427.662 38797.901 37799.972 38022.889 Direct efficiency 2% 6% 2% 3% Direct velocity 1432.174 845.464 1379.807 1317.446 Percentage direct scans 3% 2% 3% 3% Page writes by reclaim 0 0 0 0 Page writes file 0 0 0 0 Page writes anon 0 0 0 0 Page reclaim immediate 0 0 0 1218 Page rescued immediate 0 0 0 0 Slabs scanned 15360 16384 13312 16384 Direct inode steals 0 0 0 0 Kswapd inode steals 4340 4327 1630 4323 FTrace Reclaim Statistics: congestion_wait Direct number congest waited 0 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 900 870 754 789 Direct time conditional waited 0ms 0ms 0ms 20ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 2106 2308 2116 1915 KSwapd time congest waited 139924ms 157832ms 125652ms 132516ms KSwapd full congest waited 1346 1530 1202 1278 KSwapd number conditional waited 12922 16320 10943 14670 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Reclaim statistics are not radically changed. The stall times in kswapd are massive but it is clear that it is due to calls to congestion_wait() and that is almost certainly the call in balance_pgdat(). Otherwise stalls due to dirty pages are non-existant. I ran a benchmark that stressed high-order allocation. This is very artifical load but was used in the past to evaluate lumpy reclaim and compaction. Generally I look at allocation success rates and latency figures. STRESS-HIGHALLOC 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Pass 1 81.00 ( 0.00%) 28.00 (-53.00%) 24.00 (-57.00%) 28.00 (-53.00%) Pass 2 82.00 ( 0.00%) 39.00 (-43.00%) 38.00 (-44.00%) 43.00 (-39.00%) while Rested 88.00 ( 0.00%) 87.00 (-1.00%) 88.00 ( 0.00%) 88.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 740.93 681.42 685.14 684.87 User+Sys Time Running Test (seconds) 2922.65 3269.52 3281.35 3279.44 Total Elapsed Time (seconds) 1161.73 1152.49 1159.55 1161.44 MMTests Statistics: vmstat Page Ins 4486020 2807256 2855944 2876244 Page Outs 7261600 7973688 7975320 7986120 Swap Ins 31694 0 0 0 Swap Outs 98179 0 0 0 Direct pages scanned 53494 57731 34406 113015 Kswapd pages scanned 6271173 1287481 1278174 1219095 Kswapd pages reclaimed 2029240 1281025 1260708 1201583 Direct pages reclaimed 1468 14564 16649 92456 Kswapd efficiency 32% 99% 98% 98% Kswapd velocity 5398.133 1117.130 1102.302 1049.641 Direct efficiency 2% 25% 48% 81% Direct velocity 46.047 50.092 29.672 97.306 Percentage direct scans 0% 4% 2% 8% Page writes by reclaim 1616049 0 0 0 Page writes file 1517870 0 0 0 Page writes anon 98179 0 0 0 Page reclaim immediate 103778 27339 9796 17831 Page rescued immediate 0 0 0 0 Slabs scanned 1096704 986112 980992 998400 Direct inode steals 223 215040 216736 247881 Kswapd inode steals 175331 61548 68444 63066 Kswapd skipped wait 21991 0 1 0 THP fault alloc 1 135 125 134 THP collapse alloc 393 311 228 236 THP splits 25 13 7 8 THP fault fallback 0 0 0 0 THP collapse fail 3 5 7 7 Compaction stalls 865 1270 1422 1518 Compaction success 370 401 353 383 Compaction failures 495 869 1069 1135 Compaction pages moved 870155 3828868 4036106 4423626 Compaction move failure 26429 23865 29742 27514 Success rates are completely hosed for 3.4-rc2 which is almost certainly due to commit fe2c2a106663 ("vmscan: reclaim at order 0 when compaction is enabled"). I expected this would happen for kswapd and impair allocation success rates (https://lkml.org/lkml/2012/1/25/166) but I did not anticipate this much a difference: 80% less scanning, 37% less reclaim by kswapd In comparison, reclaim/compaction is not aggressive and gives up easily which is the intended behaviour. hugetlbfs uses __GFP_REPEAT and would be much more aggressive about reclaim/compaction than THP allocations are. The stress test above is allocating like neither THP or hugetlbfs but is much closer to THP. Mainline is now impaired in terms of high order allocation under heavy load although I do not know to what degree as I did not test with __GFP_REPEAT. Keep this in mind for bugs related to hugepage pool resizing, THP allocation and high order atomic allocation failures from network devices. In terms of congestion throttling, I see the following for this test FTrace Reclaim Statistics: congestion_wait Direct number congest waited 3 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 957 512 1081 1075 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 36 4 3 5 KSwapd time congest waited 3148ms 400ms 300ms 500ms KSwapd full congest waited 30 4 3 5 KSwapd number conditional waited 88514 197 332 542 KSwapd time conditional waited 4980ms 0ms 0ms 0ms KSwapd full conditional waited 49 0 0 0 The "conditional waited" times are the most interesting as this is directly impacted by the number of dirty pages encountered during scan. As lumpy reclaim is no longer scanning contiguous ranges, it is finding fewer dirty pages. This brings wait times from about 5 seconds to 0. kswapd itself is still calling congestion_wait() so it'll still stall but it's a lot less. In terms of the type of IO we were doing, I see this FTrace Reclaim Statistics: mm_vmscan_writepage Direct writes anon sync 0 0 0 0 Direct writes anon async 0 0 0 0 Direct writes file sync 0 0 0 0 Direct writes file async 0 0 0 0 Direct writes mixed sync 0 0 0 0 Direct writes mixed async 0 0 0 0 KSwapd writes anon sync 0 0 0 0 KSwapd writes anon async 91682 0 0 0 KSwapd writes file sync 0 0 0 0 KSwapd writes file async 822629 0 0 0 KSwapd writes mixed sync 0 0 0 0 KSwapd writes mixed async 0 0 0 0 In 3.2, kswapd was doing a bunch of async writes of pages but reclaim/compaction was never reaching a point where it was doing sync IO. This does not guarantee that reclaim/compaction was not calling wait_on_page_writeback() but I would consider it unlikely. It indicates that merging patches 2 and 3 to stop reclaim/compaction calling wait_on_page_writeback() should be safe. This patch: Lumpy reclaim had a purpose but in the mind of some, it was to kick the system so hard it trashed. For others the purpose was to complicate vmscan.c. Over time it was giving softer shoes and a nicer attitude but memory compaction needs to step up and replace it so this patch sends lumpy reclaim to the farm. The tracepoint format changes for isolating LRU pages with this patch applied. Furthermore reclaim/compaction can no longer queue dirty pages in pageout() if the underlying BDI is congested. Lumpy reclaim used this logic and reclaim/compaction was using it in error. Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Cc: Hugh Dickins <hughd@google.com> Cc: Ying Han <yinghan@google.com> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-30 06:06:19 +08:00
TP_ARGS(order, nr_requested, nr_scanned, nr_taken, isolate_mode, file)
);
DEFINE_EVENT(mm_vmscan_lru_isolate_template, mm_vmscan_memcg_isolate,
TP_PROTO(int order,
unsigned long nr_requested,
unsigned long nr_scanned,
unsigned long nr_taken,
isolate_mode_t isolate_mode,
int file),
mm: vmscan: remove lumpy reclaim This series removes lumpy reclaim and some stalling logic that was unintentionally being used by memory compaction. The end result is that stalling on dirty pages during page reclaim now depends on wait_iff_congested(). Four kernels were compared 3.3.0 vanilla 3.4.0-rc2 vanilla 3.4.0-rc2 lumpyremove-v2 is patch one from this series 3.4.0-rc2 nosync-v2r3 is the full series Removing lumpy reclaim saves almost 900 bytes of text whereas the full series removes 1200 bytes. text data bss dec hex filename 6740375 1927944 2260992 10929311 a6c49f vmlinux-3.4.0-rc2-vanilla 6739479 1927944 2260992 10928415 a6c11f vmlinux-3.4.0-rc2-lumpyremove-v2 6739159 1927944 2260992 10928095 a6bfdf vmlinux-3.4.0-rc2-nosync-v2 There are behaviour changes in the series and so tests were run with monitoring of ftrace events. This disrupts results so the performance results are distorted but the new behaviour should be clearer. fs-mark running in a threaded configuration showed little of interest as it did not push reclaim aggressively FS-Mark Multi Threaded 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Files/s min 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s mean 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Files/s stddev 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Files/s max 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) 3.20 ( 0.00%) Overhead min 508667.00 ( 0.00%) 521350.00 (-2.49%) 544292.00 (-7.00%) 547168.00 (-7.57%) Overhead mean 551185.00 ( 0.00%) 652690.73 (-18.42%) 991208.40 (-79.83%) 570130.53 (-3.44%) Overhead stddev 18200.69 ( 0.00%) 331958.29 (-1723.88%) 1579579.43 (-8578.68%) 9576.81 (47.38%) Overhead max 576775.00 ( 0.00%) 1846634.00 (-220.17%) 6901055.00 (-1096.49%) 585675.00 (-1.54%) MMTests Statistics: duration Sys Time Running Test (seconds) 309.90 300.95 307.33 298.95 User+Sys Time Running Test (seconds) 319.32 309.67 315.69 307.51 Total Elapsed Time (seconds) 1187.85 1193.09 1191.98 1193.73 MMTests Statistics: vmstat Page Ins 80532 82212 81420 79480 Page Outs 111434984 111456240 111437376 111582628 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 44881 27889 27453 34843 Kswapd pages scanned 25841428 25860774 25861233 25843212 Kswapd pages reclaimed 25841393 25860741 25861199 25843179 Direct pages reclaimed 44881 27889 27453 34843 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 21754.791 21675.460 21696.029 21649.127 Direct efficiency 100% 100% 100% 100% Direct velocity 37.783 23.375 23.031 29.188 Percentage direct scans 0% 0% 0% 0% ftrace showed that there was no stalling on writeback or pages submitted for IO from reclaim context. postmark was similar and while it was more interesting, it also did not push reclaim heavily. POSTMARK 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Transactions per second: 16.00 ( 0.00%) 20.00 (25.00%) 18.00 (12.50%) 17.00 ( 6.25%) Data megabytes read per second: 18.80 ( 0.00%) 24.27 (29.10%) 22.26 (18.40%) 20.54 ( 9.26%) Data megabytes written per second: 35.83 ( 0.00%) 46.25 (29.08%) 42.42 (18.39%) 39.14 ( 9.24%) Files created alone per second: 28.00 ( 0.00%) 38.00 (35.71%) 34.00 (21.43%) 30.00 ( 7.14%) Files create/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) Files deleted alone per second: 556.00 ( 0.00%) 1224.00 (120.14%) 3062.00 (450.72%) 6124.00 (1001.44%) Files delete/transact per second: 8.00 ( 0.00%) 10.00 (25.00%) 9.00 (12.50%) 8.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 113.34 107.99 109.73 108.72 User+Sys Time Running Test (seconds) 145.51 139.81 143.32 143.55 Total Elapsed Time (seconds) 1159.16 899.23 980.17 1062.27 MMTests Statistics: vmstat Page Ins 13710192 13729032 13727944 13760136 Page Outs 43071140 42987228 42733684 42931624 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 0 0 0 0 Kswapd pages scanned 9941613 9937443 9939085 9929154 Kswapd pages reclaimed 9940926 9936751 9938397 9928465 Direct pages reclaimed 0 0 0 0 Kswapd efficiency 99% 99% 99% 99% Kswapd velocity 8576.567 11051.058 10140.164 9347.109 Direct efficiency 100% 100% 100% 100% Direct velocity 0.000 0.000 0.000 0.000 It looks like here that the full series regresses performance but as ftrace showed no usage of wait_iff_congested() or sync reclaim I am assuming it's a disruption due to monitoring. Other data such as memory usage, page IO, swap IO all looked similar. Running a benchmark with a plain DD showed nothing very interesting. The full series stalled in wait_iff_congested() slightly less but stall times on vanilla kernels were marginal. Running a benchmark that hammered on file-backed mappings showed stalls due to congestion but not in sync writebacks MICRO 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 MMTests Statistics: duration Sys Time Running Test (seconds) 308.13 294.50 298.75 299.53 User+Sys Time Running Test (seconds) 330.45 316.28 318.93 320.79 Total Elapsed Time (seconds) 1814.90 1833.88 1821.14 1832.91 MMTests Statistics: vmstat Page Ins 108712 120708 97224 110344 Page Outs 155514576 156017404 155813676 156193256 Swap Ins 0 0 0 0 Swap Outs 0 0 0 0 Direct pages scanned 2599253 1550480 2512822 2414760 Kswapd pages scanned 69742364 71150694 68839041 69692533 Kswapd pages reclaimed 34824488 34773341 34796602 34799396 Direct pages reclaimed 53693 94750 61792 75205 Kswapd efficiency 49% 48% 50% 49% Kswapd velocity 38427.662 38797.901 37799.972 38022.889 Direct efficiency 2% 6% 2% 3% Direct velocity 1432.174 845.464 1379.807 1317.446 Percentage direct scans 3% 2% 3% 3% Page writes by reclaim 0 0 0 0 Page writes file 0 0 0 0 Page writes anon 0 0 0 0 Page reclaim immediate 0 0 0 1218 Page rescued immediate 0 0 0 0 Slabs scanned 15360 16384 13312 16384 Direct inode steals 0 0 0 0 Kswapd inode steals 4340 4327 1630 4323 FTrace Reclaim Statistics: congestion_wait Direct number congest waited 0 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 900 870 754 789 Direct time conditional waited 0ms 0ms 0ms 20ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 2106 2308 2116 1915 KSwapd time congest waited 139924ms 157832ms 125652ms 132516ms KSwapd full congest waited 1346 1530 1202 1278 KSwapd number conditional waited 12922 16320 10943 14670 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Reclaim statistics are not radically changed. The stall times in kswapd are massive but it is clear that it is due to calls to congestion_wait() and that is almost certainly the call in balance_pgdat(). Otherwise stalls due to dirty pages are non-existant. I ran a benchmark that stressed high-order allocation. This is very artifical load but was used in the past to evaluate lumpy reclaim and compaction. Generally I look at allocation success rates and latency figures. STRESS-HIGHALLOC 3.3.0-vanilla rc2-vanilla lumpyremove-v2r3 nosync-v2r3 Pass 1 81.00 ( 0.00%) 28.00 (-53.00%) 24.00 (-57.00%) 28.00 (-53.00%) Pass 2 82.00 ( 0.00%) 39.00 (-43.00%) 38.00 (-44.00%) 43.00 (-39.00%) while Rested 88.00 ( 0.00%) 87.00 (-1.00%) 88.00 ( 0.00%) 88.00 ( 0.00%) MMTests Statistics: duration Sys Time Running Test (seconds) 740.93 681.42 685.14 684.87 User+Sys Time Running Test (seconds) 2922.65 3269.52 3281.35 3279.44 Total Elapsed Time (seconds) 1161.73 1152.49 1159.55 1161.44 MMTests Statistics: vmstat Page Ins 4486020 2807256 2855944 2876244 Page Outs 7261600 7973688 7975320 7986120 Swap Ins 31694 0 0 0 Swap Outs 98179 0 0 0 Direct pages scanned 53494 57731 34406 113015 Kswapd pages scanned 6271173 1287481 1278174 1219095 Kswapd pages reclaimed 2029240 1281025 1260708 1201583 Direct pages reclaimed 1468 14564 16649 92456 Kswapd efficiency 32% 99% 98% 98% Kswapd velocity 5398.133 1117.130 1102.302 1049.641 Direct efficiency 2% 25% 48% 81% Direct velocity 46.047 50.092 29.672 97.306 Percentage direct scans 0% 4% 2% 8% Page writes by reclaim 1616049 0 0 0 Page writes file 1517870 0 0 0 Page writes anon 98179 0 0 0 Page reclaim immediate 103778 27339 9796 17831 Page rescued immediate 0 0 0 0 Slabs scanned 1096704 986112 980992 998400 Direct inode steals 223 215040 216736 247881 Kswapd inode steals 175331 61548 68444 63066 Kswapd skipped wait 21991 0 1 0 THP fault alloc 1 135 125 134 THP collapse alloc 393 311 228 236 THP splits 25 13 7 8 THP fault fallback 0 0 0 0 THP collapse fail 3 5 7 7 Compaction stalls 865 1270 1422 1518 Compaction success 370 401 353 383 Compaction failures 495 869 1069 1135 Compaction pages moved 870155 3828868 4036106 4423626 Compaction move failure 26429 23865 29742 27514 Success rates are completely hosed for 3.4-rc2 which is almost certainly due to commit fe2c2a106663 ("vmscan: reclaim at order 0 when compaction is enabled"). I expected this would happen for kswapd and impair allocation success rates (https://lkml.org/lkml/2012/1/25/166) but I did not anticipate this much a difference: 80% less scanning, 37% less reclaim by kswapd In comparison, reclaim/compaction is not aggressive and gives up easily which is the intended behaviour. hugetlbfs uses __GFP_REPEAT and would be much more aggressive about reclaim/compaction than THP allocations are. The stress test above is allocating like neither THP or hugetlbfs but is much closer to THP. Mainline is now impaired in terms of high order allocation under heavy load although I do not know to what degree as I did not test with __GFP_REPEAT. Keep this in mind for bugs related to hugepage pool resizing, THP allocation and high order atomic allocation failures from network devices. In terms of congestion throttling, I see the following for this test FTrace Reclaim Statistics: congestion_wait Direct number congest waited 3 0 0 0 Direct time congest waited 0ms 0ms 0ms 0ms Direct full congest waited 0 0 0 0 Direct number conditional waited 957 512 1081 1075 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 36 4 3 5 KSwapd time congest waited 3148ms 400ms 300ms 500ms KSwapd full congest waited 30 4 3 5 KSwapd number conditional waited 88514 197 332 542 KSwapd time conditional waited 4980ms 0ms 0ms 0ms KSwapd full conditional waited 49 0 0 0 The "conditional waited" times are the most interesting as this is directly impacted by the number of dirty pages encountered during scan. As lumpy reclaim is no longer scanning contiguous ranges, it is finding fewer dirty pages. This brings wait times from about 5 seconds to 0. kswapd itself is still calling congestion_wait() so it'll still stall but it's a lot less. In terms of the type of IO we were doing, I see this FTrace Reclaim Statistics: mm_vmscan_writepage Direct writes anon sync 0 0 0 0 Direct writes anon async 0 0 0 0 Direct writes file sync 0 0 0 0 Direct writes file async 0 0 0 0 Direct writes mixed sync 0 0 0 0 Direct writes mixed async 0 0 0 0 KSwapd writes anon sync 0 0 0 0 KSwapd writes anon async 91682 0 0 0 KSwapd writes file sync 0 0 0 0 KSwapd writes file async 822629 0 0 0 KSwapd writes mixed sync 0 0 0 0 KSwapd writes mixed async 0 0 0 0 In 3.2, kswapd was doing a bunch of async writes of pages but reclaim/compaction was never reaching a point where it was doing sync IO. This does not guarantee that reclaim/compaction was not calling wait_on_page_writeback() but I would consider it unlikely. It indicates that merging patches 2 and 3 to stop reclaim/compaction calling wait_on_page_writeback() should be safe. This patch: Lumpy reclaim had a purpose but in the mind of some, it was to kick the system so hard it trashed. For others the purpose was to complicate vmscan.c. Over time it was giving softer shoes and a nicer attitude but memory compaction needs to step up and replace it so this patch sends lumpy reclaim to the farm. The tracepoint format changes for isolating LRU pages with this patch applied. Furthermore reclaim/compaction can no longer queue dirty pages in pageout() if the underlying BDI is congested. Lumpy reclaim used this logic and reclaim/compaction was using it in error. Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Konstantin Khlebnikov <khlebnikov@openvz.org> Cc: Hugh Dickins <hughd@google.com> Cc: Ying Han <yinghan@google.com> Cc: Andy Whitcroft <apw@shadowen.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-30 06:06:19 +08:00
TP_ARGS(order, nr_requested, nr_scanned, nr_taken, isolate_mode, file)
);
TRACE_EVENT(mm_vmscan_writepage,
TP_PROTO(struct page *page,
int reclaim_flags),
TP_ARGS(page, reclaim_flags),
TP_STRUCT__entry(
__field(unsigned long, pfn)
__field(int, reclaim_flags)
),
TP_fast_assign(
__entry->pfn = page_to_pfn(page);
__entry->reclaim_flags = reclaim_flags;
),
TP_printk("page=%p pfn=%lu flags=%s",
pfn_to_page(__entry->pfn),
__entry->pfn,
show_reclaim_flags(__entry->reclaim_flags))
);
tracing, vmscan: add trace events for LRU list shrinking There have been numerous reports of stalls that pointed at the problem being somewhere in the VM. There are multiple roots to the problems which means dealing with any of the root problems in isolation is tricky to justify on their own and they would still need integration testing. This patch series puts together two different patch sets which in combination should tackle some of the root causes of latency problems being reported. Patch 1 adds a tracepoint for shrink_inactive_list. For this series, the most important results is being able to calculate the scanning/reclaim ratio as a measure of the amount of work being done by page reclaim. Patch 2 accounts for time spent in congestion_wait. Patches 3-6 were originally developed by Kosaki Motohiro but reworked for this series. It has been noted that lumpy reclaim is far too aggressive and trashes the system somewhat. As SLUB uses high-order allocations, a large cost incurred by lumpy reclaim will be noticeable. It was also reported during transparent hugepage support testing that lumpy reclaim was trashing the system and these patches should mitigate that problem without disabling lumpy reclaim. Patch 7 adds wait_iff_congested() and replaces some callers of congestion_wait(). wait_iff_congested() only sleeps if there is a BDI that is currently congested. Patch 8 notes that any BDI being congested is not necessarily a problem because there could be multiple BDIs of varying speeds and numberous zones. It attempts to track when a zone being reclaimed contains many pages backed by a congested BDI and if so, reclaimers wait on the congestion queue. I ran a number of tests with monitoring on X86, X86-64 and PPC64. Each machine had 3G of RAM and the CPUs were X86: Intel P4 2-core X86-64: AMD Phenom 4-core PPC64: PPC970MP Each used a single disk and the onboard IO controller. Dirty ratio was left at 20. I'm just going to report for X86-64 and PPC64 in a vague attempt to keep this report short. Four kernels were tested each based on v2.6.36-rc4 traceonly-v2r2: Patches 1 and 2 to instrument vmscan reclaims and congestion_wait lowlumpy-v2r3: Patches 1-6 to test if lumpy reclaim is better waitcongest-v2r3: Patches 1-7 to only wait on congestion waitwriteback-v2r4: Patches 1-8 to detect when a zone is congested nocongest-v1r5: Patches 1-3 for testing wait_iff_congestion nodirect-v1r5: Patches 1-10 to disable filesystem writeback for better IO The tests run were as follows kernbench compile-based benchmark. Smoke test performance sysbench OLTP read-only benchmark. Will be re-run in the future as read-write micro-mapped-file-stream This is a micro-benchmark from Johannes Weiner that accesses a large sparse-file through mmap(). It was configured to run in only single-CPU mode but can be indicative of how well page reclaim identifies suitable pages. stress-highalloc Tries to allocate huge pages under heavy load. kernbench, iozone and sysbench did not report any performance regression on any machine. sysbench did pressure the system lightly and there was reclaim activity but there were no difference of major interest between the kernels. X86-64 micro-mapped-file-stream traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 pgalloc_dma 1639.00 ( 0.00%) 667.00 (-145.73%) 1167.00 ( -40.45%) 578.00 (-183.56%) pgalloc_dma32 2842410.00 ( 0.00%) 2842626.00 ( 0.01%) 2843043.00 ( 0.02%) 2843014.00 ( 0.02%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 729.00 ( 0.00%) 85.00 (-757.65%) 609.00 ( -19.70%) 125.00 (-483.20%) pgsteal_dma32 2338721.00 ( 0.00%) 2447354.00 ( 4.44%) 2429536.00 ( 3.74%) 2436772.00 ( 4.02%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 1469.00 ( 0.00%) 532.00 (-176.13%) 1078.00 ( -36.27%) 220.00 (-567.73%) pgscan_kswapd_dma32 4597713.00 ( 0.00%) 4503597.00 ( -2.09%) 4295673.00 ( -7.03%) 3891686.00 ( -18.14%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 71.00 ( 0.00%) 134.00 ( 47.01%) 243.00 ( 70.78%) 352.00 ( 79.83%) pgscan_direct_dma32 305820.00 ( 0.00%) 280204.00 ( -9.14%) 600518.00 ( 49.07%) 957485.00 ( 68.06%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 16296.00 ( 0.00%) 21254.00 ( 23.33%) 18447.00 ( 11.66%) 20067.00 ( 18.79%) allocstall 443.00 ( 0.00%) 273.00 ( -62.27%) 513.00 ( 13.65%) 1568.00 ( 71.75%) These are based on the raw figures taken from /proc/vmstat. It's a rough measure of reclaim activity. Note that allocstall counts are higher because we are entering direct reclaim more often as a result of not sleeping in congestion. In itself, it's not necessarily a bad thing. It's easier to get a view of what happened from the vmscan tracepoint report. FTrace Reclaim Statistics: vmscan traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4 Direct reclaims 443 273 513 1568 Direct reclaim pages scanned 305968 280402 600825 957933 Direct reclaim pages reclaimed 43503 19005 30327 117191 Direct reclaim write file async I/O 0 0 0 0 Direct reclaim write anon async I/O 0 3 4 12 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 187649 132338 191695 267701 Kswapd wakeups 3 1 4 1 Kswapd pages scanned 4599269 4454162 4296815 3891906 Kswapd pages reclaimed 2295947 2428434 2399818 2319706 Kswapd reclaim write file async I/O 1 0 1 1 Kswapd reclaim write anon async I/O 59 187 41 222 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4.34 2.52 6.63 2.96 Time kswapd awake (seconds) 11.15 10.25 11.01 10.19 Total pages scanned 4905237 4734564 4897640 4849839 Total pages reclaimed 2339450 2447439 2430145 2436897 %age total pages scanned/reclaimed 47.69% 51.69% 49.62% 50.25% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 29.23% 19.02% 38.48% 20.25% Percentage Time kswapd Awake 78.58% 78.85% 76.83% 79.86% What is interesting here for nocongest in particular is that while direct reclaim scans more pages, the overall number of pages scanned remains the same and the ratio of pages scanned to pages reclaimed is more or less the same. In other words, while we are sleeping less, reclaim is not doing more work and as direct reclaim and kswapd is awake for less time, it would appear to be doing less work. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 87 196 64 0 Direct time congest waited 4604ms 4732ms 5420ms 0ms Direct full congest waited 72 145 53 0 Direct number conditional waited 0 0 324 1315 Direct time conditional waited 0ms 0ms 0ms 0ms Direct full conditional waited 0 0 0 0 KSwapd number congest waited 20 10 15 7 KSwapd time congest waited 1264ms 536ms 884ms 284ms KSwapd full congest waited 10 4 6 2 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 8 seconds asleep in direct reclaim and no time at all asleep with the patches. MMTests Statistics: duration User/Sys Time Running Test (seconds) 10.51 10.73 10.6 11.66 Total Elapsed Time (seconds) 14.19 13.00 14.33 12.76 Overall, the tests completed faster. It is interesting to note that backing off further when a zone is congested and not just a BDI was more efficient overall. PPC64 micro-mapped-file-stream pgalloc_dma 3024660.00 ( 0.00%) 3027185.00 ( 0.08%) 3025845.00 ( 0.04%) 3026281.00 ( 0.05%) pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgsteal_dma 2508073.00 ( 0.00%) 2565351.00 ( 2.23%) 2463577.00 ( -1.81%) 2532263.00 ( 0.96%) pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_kswapd_dma 4601307.00 ( 0.00%) 4128076.00 ( -11.46%) 3912317.00 ( -17.61%) 3377165.00 ( -36.25%) pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pgscan_direct_dma 629825.00 ( 0.00%) 971622.00 ( 35.18%) 1063938.00 ( 40.80%) 1711935.00 ( 63.21%) pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) pageoutrun 27776.00 ( 0.00%) 20458.00 ( -35.77%) 18763.00 ( -48.04%) 18157.00 ( -52.98%) allocstall 977.00 ( 0.00%) 2751.00 ( 64.49%) 2098.00 ( 53.43%) 5136.00 ( 80.98%) Similar trends to x86-64. allocstalls are up but it's not necessarily bad. FTrace Reclaim Statistics: vmscan Direct reclaims 977 2709 2098 5136 Direct reclaim pages scanned 629825 963814 1063938 1711935 Direct reclaim pages reclaimed 75550 242538 150904 387647 Direct reclaim write file async I/O 0 0 0 2 Direct reclaim write anon async I/O 0 10 0 4 Direct reclaim write file sync I/O 0 0 0 0 Direct reclaim write anon sync I/O 0 0 0 0 Wake kswapd requests 392119 1201712 571935 571921 Kswapd wakeups 3 2 3 3 Kswapd pages scanned 4601307 4128076 3912317 3377165 Kswapd pages reclaimed 2432523 2318797 2312673 2144616 Kswapd reclaim write file async I/O 20 1 1 1 Kswapd reclaim write anon async I/O 57 132 11 121 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 6.19 7.30 13.04 10.88 Time kswapd awake (seconds) 21.73 26.51 25.55 23.90 Total pages scanned 5231132 5091890 4976255 5089100 Total pages reclaimed 2508073 2561335 2463577 2532263 %age total pages scanned/reclaimed 47.95% 50.30% 49.51% 49.76% %age total pages scanned/written 0.00% 0.00% 0.00% 0.00% %age file pages scanned/written 0.00% 0.00% 0.00% 0.00% Percentage Time Spent Direct Reclaim 18.89% 20.65% 32.65% 27.65% Percentage Time kswapd Awake 72.39% 80.68% 78.21% 77.40% Again, a similar trend that the congestion_wait changes mean that direct reclaim scans more pages but the overall number of pages scanned while slightly reduced, are very similar. The ratio of scanning/reclaimed remains roughly similar. The downside is that kswapd and direct reclaim was awake longer and for a larger percentage of the overall workload. It's possible there were big differences in the amount of time spent reclaiming slab pages between the different kernels which is plausible considering that the micro tests runs after fsmark and sysbench. Trace Reclaim Statistics: congestion_wait Direct number congest waited 845 1312 104 0 Direct time congest waited 19416ms 26560ms 7544ms 0ms Direct full congest waited 745 1105 72 0 Direct number conditional waited 0 0 1322 2935 Direct time conditional waited 0ms 0ms 12ms 312ms Direct full conditional waited 0 0 0 3 KSwapd number congest waited 39 102 75 63 KSwapd time congest waited 2484ms 6760ms 5756ms 3716ms KSwapd full congest waited 20 48 46 25 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 The vanilla kernel spent 20 seconds asleep in direct reclaim and only 312ms asleep with the patches. The time kswapd spent congest waited was also reduced by a large factor. MMTests Statistics: duration ser/Sys Time Running Test (seconds) 26.58 28.05 26.9 28.47 Total Elapsed Time (seconds) 30.02 32.86 32.67 30.88 With all patches applies, the completion times are very similar. X86-64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 82.00 ( 0.00%) 84.00 ( 2.00%) 85.00 ( 3.00%) 85.00 ( 3.00%) Pass 2 90.00 ( 0.00%) 87.00 (-3.00%) 88.00 (-2.00%) 89.00 (-1.00%) At Rest 92.00 ( 0.00%) 90.00 (-2.00%) 90.00 (-2.00%) 91.00 (-1.00%) Success figures across the board are broadly similar. traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 1045 944 886 887 Direct reclaim pages scanned 135091 119604 109382 101019 Direct reclaim pages reclaimed 88599 47535 47863 46671 Direct reclaim write file async I/O 494 283 465 280 Direct reclaim write anon async I/O 29357 13710 16656 13462 Direct reclaim write file sync I/O 154 2 2 3 Direct reclaim write anon sync I/O 14594 571 509 561 Wake kswapd requests 7491 933 872 892 Kswapd wakeups 814 778 731 780 Kswapd pages scanned 7290822 15341158 11916436 13703442 Kswapd pages reclaimed 3587336 3142496 3094392 3187151 Kswapd reclaim write file async I/O 91975 32317 28022 29628 Kswapd reclaim write anon async I/O 1992022 789307 829745 849769 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 4588.93 2467.16 2495.41 2547.07 Time kswapd awake (seconds) 2497.66 1020.16 1098.06 1176.82 Total pages scanned 7425913 15460762 12025818 13804461 Total pages reclaimed 3675935 3190031 3142255 3233822 %age total pages scanned/reclaimed 49.50% 20.63% 26.13% 23.43% %age total pages scanned/written 28.66% 5.41% 7.28% 6.47% %age file pages scanned/written 1.25% 0.21% 0.24% 0.22% Percentage Time Spent Direct Reclaim 57.33% 42.15% 42.41% 42.99% Percentage Time kswapd Awake 43.56% 27.87% 29.76% 31.25% Scanned/reclaimed ratios again look good with big improvements in efficiency. The Scanned/written ratios also look much improved. With a better scanned/written ration, there is an expectation that IO would be more efficient and indeed, the time spent in direct reclaim is much reduced by the full series and kswapd spends a little less time awake. Overall, indications here are that allocations were happening much faster and this can be seen with a graph of the latency figures as the allocations were taking place http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-hydra-mean.ps FTrace Reclaim Statistics: congestion_wait Direct number congest waited 1333 204 169 4 Direct time congest waited 78896ms 8288ms 7260ms 200ms Direct full congest waited 756 92 69 2 Direct number conditional waited 0 0 26 186 Direct time conditional waited 0ms 0ms 0ms 2504ms Direct full conditional waited 0 0 0 25 KSwapd number congest waited 4 395 227 282 KSwapd time congest waited 384ms 25136ms 10508ms 18380ms KSwapd full congest waited 3 232 98 176 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 KSwapd full conditional waited 318 0 312 9 Overall, the time spent speeping is reduced. kswapd is still hitting congestion_wait() but that is because there are callers remaining where it wasn't clear in advance if they should be changed to wait_iff_congested() or not. Overall the sleep imes are reduced though - from 79ish seconds to about 19. MMTests Statistics: duration User/Sys Time Running Test (seconds) 3415.43 3386.65 3388.39 3377.5 Total Elapsed Time (seconds) 5733.48 3660.33 3689.41 3765.39 With the full series, the time to complete the tests are reduced by 30% PPC64 STRESS-HIGHALLOC traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Pass 1 17.00 ( 0.00%) 34.00 (17.00%) 38.00 (21.00%) 43.00 (26.00%) Pass 2 25.00 ( 0.00%) 37.00 (12.00%) 42.00 (17.00%) 46.00 (21.00%) At Rest 49.00 ( 0.00%) 43.00 (-6.00%) 45.00 (-4.00%) 51.00 ( 2.00%) Success rates there are *way* up particularly considering that the 16MB huge pages on PPC64 mean that it's always much harder to allocate them. FTrace Reclaim Statistics: vmscan stress-highalloc stress-highalloc stress-highalloc stress-highalloc traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4 Direct reclaims 499 505 564 509 Direct reclaim pages scanned 223478 41898 51818 45605 Direct reclaim pages reclaimed 137730 21148 27161 23455 Direct reclaim write file async I/O 399 136 162 136 Direct reclaim write anon async I/O 46977 2865 4686 3998 Direct reclaim write file sync I/O 29 0 1 3 Direct reclaim write anon sync I/O 31023 159 237 239 Wake kswapd requests 420 351 360 326 Kswapd wakeups 185 294 249 277 Kswapd pages scanned 15703488 16392500 17821724 17598737 Kswapd pages reclaimed 5808466 2908858 3139386 3145435 Kswapd reclaim write file async I/O 159938 18400 18717 13473 Kswapd reclaim write anon async I/O 3467554 228957 322799 234278 Kswapd reclaim write file sync I/O 0 0 0 0 Kswapd reclaim write anon sync I/O 0 0 0 0 Time stalled direct reclaim (seconds) 9665.35 1707.81 2374.32 1871.23 Time kswapd awake (seconds) 9401.21 1367.86 1951.75 1328.88 Total pages scanned 15926966 16434398 17873542 17644342 Total pages reclaimed 5946196 2930006 3166547 3168890 %age total pages scanned/reclaimed 37.33% 17.83% 17.72% 17.96% %age total pages scanned/written 23.27% 1.52% 1.94% 1.43% %age file pages scanned/written 1.01% 0.11% 0.11% 0.08% Percentage Time Spent Direct Reclaim 44.55% 35.10% 41.42% 36.91% Percentage Time kswapd Awake 86.71% 43.58% 52.67% 41.14% While the scanning rates are slightly up, the scanned/reclaimed and scanned/written figures are much improved. The time spent in direct reclaim and with kswapd are massively reduced, mostly by the lowlumpy patches. FTrace Reclaim Statistics: congestion_wait Direct number congest waited 725 303 126 3 Direct time congest waited 45524ms 9180ms 5936ms 300ms Direct full congest waited 487 190 52 3 Direct number conditional waited 0 0 200 301 Direct time conditional waited 0ms 0ms 0ms 1904ms Direct full conditional waited 0 0 0 19 KSwapd number congest waited 0 2 23 4 KSwapd time congest waited 0ms 200ms 420ms 404ms KSwapd full congest waited 0 2 2 4 KSwapd number conditional waited 0 0 0 0 KSwapd time conditional waited 0ms 0ms 0ms 0ms KSwapd full conditional waited 0 0 0 0 Not as dramatic a story here but the time spent asleep is reduced and we can still see what wait_iff_congested is going to sleep when necessary. MMTests Statistics: duration User/Sys Time Running Test (seconds) 12028.09 3157.17 3357.79 3199.16 Total Elapsed Time (seconds) 10842.07 3138.72 3705.54 3229.85 The time to complete this test goes way down. With the full series, we are allocating over twice the number of huge pages in 30% of the time and there is a corresponding impact on the allocation latency graph available at. http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-powyah-mean.ps This patch: Add a trace event for shrink_inactive_list() and updates the sample postprocessing script appropriately. It can be used to determine how many pages were reclaimed and for non-lumpy reclaim where exactly the pages were reclaimed from. Signed-off-by: Mel Gorman <mel@csn.ul.ie> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-27 05:21:40 +08:00
TRACE_EVENT(mm_vmscan_lru_shrink_inactive,
TP_PROTO(int nid, int zid,
unsigned long nr_scanned, unsigned long nr_reclaimed,
int priority, int reclaim_flags),
TP_ARGS(nid, zid, nr_scanned, nr_reclaimed, priority, reclaim_flags),
TP_STRUCT__entry(
__field(int, nid)
__field(int, zid)
__field(unsigned long, nr_scanned)
__field(unsigned long, nr_reclaimed)
__field(int, priority)
__field(int, reclaim_flags)
),
TP_fast_assign(
__entry->nid = nid;
__entry->zid = zid;
__entry->nr_scanned = nr_scanned;
__entry->nr_reclaimed = nr_reclaimed;
__entry->priority = priority;
__entry->reclaim_flags = reclaim_flags;
),
TP_printk("nid=%d zid=%d nr_scanned=%ld nr_reclaimed=%ld priority=%d flags=%s",
__entry->nid, __entry->zid,
__entry->nr_scanned, __entry->nr_reclaimed,
__entry->priority,
show_reclaim_flags(__entry->reclaim_flags))
);
#endif /* _TRACE_VMSCAN_H */
/* This part must be outside protection */
#include <trace/define_trace.h>