From: Kairui Song <kasong@xxxxxxxxxxx> Hi, linux-mm I noticed MGLRU not working very well on certain workflows, which is observed on some instances on some heavily stressed machines. I found this was related to refault distance detection, when the file page workingset size exceeds total memory, and the access distance (the left-shift time of a page before it gets activated, considering LRU starts from right) of file pages also larger than total memory. All file pages are stuck on the oldest generation and getting read-in then evicted permutably, few get activated and stay in memory. This series tries to fix this problem by rework the refault distance detection to better fit MGLRU, and also tries to use a unified algorithm for both MGLRU and Inactive/Active LRU. Patch 1/4 reworked the refault distance detection model for Inactive/Active LRU. Patch 2/4 and 3/4 are simplification and prepare. Patch 4/4 applies the modified refault distance detection for MGLRU. Following benchmark showed 5x improvement: To simulate the workflow, I setup a 3-replicated mongodb cluster using docker, each in a standalone cgroup, set to use 5 gb of cache and 10g of oplog, on a 32G VM. The benchmark is done using https://github.com/apavlo/py-tpcc.git, modified to run STOCK_LEVEL query only, for simulating slow query and get a stable result. Before the patch (with 10G swap, the result won't change whether swap is on or not): $ tpcc.py --config=mongodb.config mongodb --duration=900 --warehouses=500 --clients=30 ================================================================== Execution Results after 904 seconds ------------------------------------------------------------------ Executed Time (µs) Rate STOCK_LEVEL 503 27150226136.4 0.02 txn/s ------------------------------------------------------------------ TOTAL 503 27150226136.4 0.02 txn/s $ cat /proc/vmstat | grep working workingset_nodes 53391 workingset_refault_anon 0 workingset_refault_file 23856735 workingset_activate_anon 0 workingset_activate_file 23845737 workingset_restore_anon 0 workingset_restore_file 18280692 workingset_nodereclaim 1024 $ free -m total used free shared buff/cache available Mem: 31837 6752 379 23 24706 24607 Swap: 10239 0 10239 After the patch (with 10G swap on same disk, similar result using ZRAM): $ tpcc.py --config=mongodb.config mongodb --duration=900 --warehouses=500 --clients=30 ================================================================== Execution Results after 903 seconds ------------------------------------------------------------------ Executed Time (µs) Rate STOCK_LEVEL 2575 27094953498.8 0.10 txn/s ------------------------------------------------------------------ TOTAL 2575 27094953498.8 0.10 txn/s $ cat /proc/vmstat | grep working workingset_nodes 78249 workingset_refault_anon 10139 workingset_refault_file 23001863 workingset_activate_anon 7238 workingset_activate_file 6718032 workingset_restore_anon 7432 workingset_restore_file 6719406 workingset_nodereclaim 9747 $ free -m total used free shared buff/cache available Mem: 31837 7376 320 3 24140 24014 Swap: 10239 1662 8577 The performance is 5x times better than before, and the idle anon pages now can get swapped out as expected. Testing with lower stress also shows a improvement. I also checked the benchmark with memtier/memcached and fio, using similar setup as in commit ac35a4902374 but scaled down to fit in my test environment: memtier test (with 16G ramdisk as swap and 2G cgroup limit): memcached -u nobody -m 16384 -s /tmp/memcached.socket -a 0766 \ -t 12 -B binary & memtier_benchmark -S /tmp/memcached.socket -P memcache_binary -n allkeys\ --key-minimum=1 --key-maximum=24000000 --key-pattern=P:P -c 1 \ -t 12 --ratio 1:0 --pipeline 8 -d 2000 -x 6 fio test (with 16G ramdisk on /mnt and 4G cgroup limit): fio -name=refault --numjobs=12 --directory=/mnt --size=1024m \ --buffered=1 --ioengine=io_uring --iodepth=128 \ --iodepth_batch_submit=32 --iodepth_batch_complete=32 \ --rw=randread --random_distribution=random --norandommap \ --time_based --ramp_time=5m --runtime=5m --group_reporting Before this patch: memcached: Ops/sec Hits/sec Misses/sec Avg. Latency p50 Latency p99 Latency p99.9 Latency KB/sec Best 52832.79 0.00 0.00 1.82042 1.70300 4.54300 6.27100 105641.69 Worst 46613.56 0.00 0.00 2.05686 1.77500 7.80700 11.83900 93206.05 Avg (6x) 51024.85 0.00 0.00 1.88506 1.73500 5.43900 9.47100 102026.64 fio: read: IOPS=2211k, BW=8637MiB/s (9056MB/s)(2530GiB/300001msec) After this patch: memcached: Ops/sec Avg. Latency p50 Latency p99 Latency p99.9 Latency KB/sec Best 54218.92 1.76930 1.65500 4.41500 6.27100 108413.34 Worst 47640.13 2.01495 1.74300 7.64700 11.64700 95258.72 Avg (6x) 51408.33 1.86988 1.71900 5.43900 9.34300 102793.42 fio: read: IOPS=2166k, BW=8462MiB/s (8873MB/s)(2479GiB/300001msec) memcached looks ok but there is a %2 performance drop for FIO test, and after some profiling this is mainly caused by the extra atomic operations and new functions, there seems to be no LRU accuracy drop. Sending this as RFC as I'm not entirely sure if this is the right way to fix this issue, of if this is a generic issue or considered more of a misconfiguration. Any suggetions about how should I test it is welcomed. Signed-off-by: Kairui Song <kasong@xxxxxxxxxxx> Kairui Song (4): workingset: simplify and use a more intuitive model workingset: simplify lru_gen_test_recent lru_gen: convert avg_total and avg_refaulted to atomic workingset, lru_gen: apply refault-distance based re-activation include/linux/mmzone.h | 4 +- include/linux/swap.h | 2 - mm/swap.c | 1 - mm/vmscan.c | 18 ++- mm/workingset.c | 315 ++++++++++++++++++++++------------------- 5 files changed, 179 insertions(+), 161 deletions(-) -- 2.41.0
