lkml.org 
[lkml]   [2020]   [Jan]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectPaging out when Free memory is low but there is Available memory (kernel 5.3 and 5.4 at least)
Date
I've observed a problem in recent kernels whereby the system will page out, continuously and seemingly needlessly, when Free memory is low but not exhausted. The system may have gigabytes of Available memory, yet begins paging out when buffer/filesystem cache squeezes out the remaining Free memory. Strangely, swap space utilisation does not increase over time, despite the continuous page-outs with little or no page-in activity.

I've observed this on 5.3 and 5.4 series kernels on Arch Linux on a Dell Optiplex 9020 (Intel Q87 Express, i5-4670 CPU, 12 GiB RAM).


To reproduce:


1. Get the system in a low-memory state, but not so low that it has to start paging, e.g. using stress/stress-ng:

stress --vm-bytes $(awk '/MemAvailable/{printf "%d\n", $2 * 0.95;}' < /proc/meminfo)k --vm-keep -m 1

Confirm that the system isn't paging, and that memory pressure reported by /proc/pressure/memory remains low:

cat /proc/pressure/memory
some avg10=0.00 avg60=0.00 avg300=1.54 total=414331152
full avg10=0.00 avg60=0.00 avg300=1.42 total=377927386


2. Perform an additional task that will fill up the filesystem cache and/or I/O buffers, such as dd.

dd if=/dev/sda of=/dev/null bs=1048576


3. Observe system behaviour (with `vmstat 1`, `xosview`, etc.) while the free memory decreases. Here's what I witness using vmstat:

procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
2 0 62464 456132 591328 484656 0 0 0 0 1643 2859 26 1 74 0 0
1 0 62464 456076 591328 484676 0 0 0 0 1684 2813 26 0 74 0 0
1 0 62464 456084 591328 484660 0 0 88 0 2156 3212 25 1 73 1 0
1 0 62208 456084 591328 485004 0 0 0 0 2284 3419 26 1 73 0 0
1 0 62208 456184 591328 485004 0 0 0 0 2027 2990 26 1 74 0 0
1 0 62208 456020 591328 485016 0 0 0 0 1960 2983 26 1 74 0 0
4 0 62208 456052 591328 485016 0 0 0 0 2034 3046 26 1 73 0 0
1 0 62208 456084 591328 485000 0 0 0 0 1882 2758 26 1 74 0 0
1 1 62208 404384 642532 485000 0 0 51196 12 2027 3149 25 2 64 9 0 <- dd begins
1 1 62344 385928 686292 460928 0 200 124928 200 2218 3785 25 3 55 18 0
1 1 70356 394056 681360 453036 0 21700 81920 21700 2373 3577 26 4 40 31 0
1 1 83576 380132 694960 439964 0 13532 99840 13532 2060 3560 25 3 48 24 0
1 1 79552 391188 685408 443776 0 4264 123904 4264 2188 3742 25 3 54 19 0
1 1 79616 395960 679076 443828 0 7628 108032 7628 2200 3686 26 2 52 20 0
1 2 83456 389928 691104 435692 0 11136 104448 11136 2203 3660 25 2 49 23 0
1 1 83968 415780 664744 435596 0 1480 121344 1480 2285 3630 25 3 54 18 0
1 1 77312 419104 662376 442284 0 1176 124416 1176 2236 3662 25 4 56 16 0
1 1 79872 398472 679452 439552 0 9164 110592 9164 2255 3703 25 3 49 23 0
1 2 83200 389140 691968 436056 0 11132 108032 11132 2113 3550 25 3 47 25 0
1 1 84480 396456 684096 435296 0 2364 109056 2364 2072 3542 26 2 51 21 0
1 3 76032 391896 698608 433428 0 6712 110664 6712 2134 3696 25 3 49 23 0
1 1 70144 408468 672888 447244 104 2016 36592 2016 2561 4609 25 2 32 41 0
2 2 77312 356556 712208 442896 0 6880 40328 7052 2223 3580 31 2 25 41 0
1 2 85504 500392 566936 449100 0 8780 30052 22128 2172 3588 26 3 19 52 0
1 2 81408 391236 675480 453360 0 2804 108716 2804 2219 3535 25 2 53 20 0
1 2 81152 379424 689692 452648 0 7408 104448 7408 2321 3665 25 3 48 23 0
1 1 85104 409144 659932 448120 0 11704 94208 11704 2286 3682 25 3 47 25 0
2 1 85368 396440 673868 448244 28 1212 99868 1288 2170 3569 25 2 46 27 0
1 2 83968 385172 683892 449572 0 6588 118784 6588 2169 3904 25 3 53 19 0
1 1 78848 388668 681108 454724 0 2172 118784 2284 2097 3724 25 3 50 22 0
1 2 70912 419592 673860 438356 0 7340 84684 7340 1942 3672 25 3 43 29 0
3 1 77824 406288 685452 433064 0 6940 55604 7060 2107 3755 25 2 34 39 0
1 2 74496 394684 697252 436316 0 4028 118784 4028 2250 3894 25 2 52 21 0
1 1 78080 417060 675348 433080 0 10884 102404 10884 2486 4609 25 3 47 25 0
1 2 73984 397144 694840 437160 48 3508 106544 3580 2039 3672 25 2 49 23 0
1 3 78080 365596 725560 434504 124 4024 31088 4024 1800 2932 25 2 27 46 0
1 3 73216 373732 716676 438908 100 1952 34952 1952 1981 3335 26 1 26 47 0
1 2 71680 380044 709484 441480 72 5204 26668 5204 1913 3376 26 1 22 51 0
1 3 71424 394544 692012 442364 48 6696 27532 6696 1909 3291 27 1 15 58 0
1 3 74908 392528 697372 438248 84 3584 27540 3600 1856 3232 25 2 15 57 0
1 2 71168 417136 662892 449732 84 3768 29044 3768 1919 3385 26 2 29 43 0
1 2 78932 417788 663360 441568 8 7932 86128 7932 2169 3579 26 2 45 27 0
1 1 85108 439408 639828 435408 0 13660 101888 13660 2157 3461 25 3 48 25 0
1 1 85364 402868 678052 434988 0 1136 123392 1136 2240 3797 25 3 56 15 0
1 1 78460 380424 700440 442888 0 1008 124928 1008 2188 3745 25 3 56 17 0
1 1 80896 386260 695628 439068 0 10108 91648 10108 2170 3572 26 2 46 27 0
1 2 84480 389552 689404 435780 0 10988 100864 10988 2015 3450 25 2 48 24 0
1 1 85404 415768 664708 434892 0 1980 122880 1980 2024 3658 25 3 53 19 0
1 2 80128 385008 696088 439256 0 2116 124928 2116 2047 3657 24 3 55 18 0
1 1 80896 390064 691756 438140 0 7784 112128 7784 2062 3621 25 3 50 21 0
1 2 77824 384716 704968 431932 0 11384 98816 11384 1968 3587 26 2 47 26 0
2 1 78848 395324 692572 431928 0 2240 118272 2240 2006 3716 25 2 54 18 0
1 1 74332 392796 694952 436552 0 3560 120832 3560 2098 3868 26 2 53 19 0
1 1 78336 402544 684296 433416 0 12032 96256 12032 1959 3421 25 3 45 27 0
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----

There is considerable paging-out activity ("so" column), despite having available memory. Swap space utilisation rises briefly, then remains roughly constant. I am surprised to see any page-outs, let alone constant. And with several kilo-blocks being paged out per second, and negligible page-in activity, shouldn't swap be growing by about a megabyte per second?

The kernel seems to think it's under some memory pressure:

cat /proc/pressure/memory
some avg10=8.01 avg60=5.44 avg300=2.70 total=422931595
full avg10=7.66 avg60=5.15 avg300=2.53 total=385972806

The paging activity stops when the dd process finishes or is killed or suspended.


4. Null test: To show that it's not the dd process that's causing the memory pressure, read from a device that won't add to buffer cache (memory hog still running):

dd if=/dev/zero of=/dev/null bs=1048576

procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
2 0 67584 432504 622944 487216 0 0 0 0 2165 3512 26 1 73 0 0
1 0 67584 432440 622944 487292 0 0 0 0 2211 3457 26 1 73 0 0
1 0 67584 432180 622944 487292 0 0 0 88 1947 3033 26 1 74 0 0
1 0 67584 431896 622944 487352 0 0 0 0 1729 2810 26 1 73 0 0
1 0 67584 431644 622944 487272 0 0 0 0 1936 2935 26 1 73 0 0
..

cat /proc/pressure/memory
some avg10=0.00 avg60=0.10 avg300=1.48 total=427081339
full avg10=0.00 avg60=0.10 avg300=1.39 total=389910888

No problem.

The problematic behaviour is also seen when the filesystem cache (rather than the buffer cache) is exercised, e.g. md5sum /some/big/files/*.

The needless paging out is also seen without a "memory hog" running alongside the buffer/fs cache exerciser - it just takes longer to reach the point at which it starts paging out, and reclaims more of the cache memory between paging bursts. A graph of memory used has a visible, irregular sawtooth shape.

Deactivating swap altogether avoids the problem. ;) Setting vm.swappiness to 1 or 0 does not.

Purging the cache with "echo 3 > /proc/sys/vm/drop_caches" stops the paging-out activity temporarily, until the buff/cache utilisation creeps up again. vmstat output again:

procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
1 0 67072 409928 21952 1107556 0 0 0 0 1772 2808 26 1 74 0 0
1 0 67072 409644 21952 1107564 0 0 0 0 2005 3320 26 1 73 0 0
1 0 67072 409644 21952 1107564 0 0 0 0 1906 3035 26 1 73 0 0
1 0 67072 415944 21952 1100744 0 0 0 0 2083 3709 27 1 72 0 0
1 0 67072 415944 21952 1100744 0 0 0 0 1960 3203 26 1 73 0 0
1 0 67072 415156 21952 1100688 0 0 0 0 1941 3307 26 1 73 0 0
1 0 67072 415604 21968 1100672 0 0 0 116 1722 2856 26 1 70 3 0
1 0 67072 415376 21968 1100672 0 0 0 0 1682 2888 26 1 73 0 0
1 1 67072 401004 35744 1100620 0 0 13780 116 1725 3006 26 1 70 3 0 <- dd begins
1 2 71936 394432 142240 994452 0 7516 106496 7564 2407 4083 27 4 49 21 0
1 1 83456 444492 231840 846960 0 19064 89600 19064 2389 4114 26 4 45 25 0
1 1 83200 470376 334240 715248 0 7496 102400 7496 2338 4018 26 5 48 21 0
1 1 83456 471480 436028 612844 0 7696 109568 7696 2390 4063 27 4 49 20 0
1 1 75116 427784 488760 611496 0 6968 111104 6968 2360 4049 26 5 51 19 0
1 1 75520 381788 534760 612608 0 7548 96256 7560 2296 3922 26 4 47 23 0
1 1 75564 387984 556520 583716 0 7080 124416 7080 2326 3986 27 5 55 14 0
1 1 75480 400808 582600 545416 0 7360 112128 7360 2357 3800 26 5 52 17 0
2 1 75520 1104992 23844 398784 0 7460 115712 7460 2449 3981 27 19 37 18 0 <- drop_caches
1 2 68608 1087920 49220 398004 100 168 26036 168 1919 3251 26 2 23 49 0
1 2 68352 1062184 72400 401876 4 0 25324 20 1875 3019 26 2 35 38 0
2 0 68352 970188 164048 402676 0 0 92364 0 2166 3662 28 3 45 24 0
1 1 68352 846924 288464 402920 0 0 124416 0 2304 3919 26 4 56 14 0
1 1 68352 723664 412880 402616 0 0 124416 0 2447 3754 26 4 55 15 0
1 1 68352 599428 537296 402492 0 0 124416 0 2467 3765 26 4 55 15 0
1 3 68352 477680 658644 402940 4 0 121376 16 2495 3886 26 3 55 15 0
1 2 75776 431236 706620 394264 0 15488 94216 15492 2410 4113 27 5 44 24 0
1 1 75636 383948 754740 394340 0 7972 114176 7972 2382 4015 26 5 49 19 0
1 1 75520 391220 747020 394252 0 7028 115712 7028 2331 3886 27 4 52 18 0
1 1 75264 393764 744824 394080 0 7360 95232 7512 2425 4030 26 4 45 24 0
2 0 70912 389928 747168 397940 0 3028 110080 3028 2427 3861 26 4 51 18 0
1 1 68352 1075872 62280 400020 0 4512 86144 4524 2325 3696 26 5 44 25 0 <- drop_caches
1 1 68352 951904 185160 401044 0 0 122880 0 2534 4059 27 5 54 14 0
1 1 68352 829684 308552 401072 0 0 123392 0 2443 3827 26 5 56 13 0
1 1 68352 696240 432968 410196 0 0 124416 0 2488 3881 28 4 53 15 0
1 1 68352 571720 557384 409932 0 0 124416 0 2444 3849 27 4 57 13 0
1 1 68352 447232 681800 410264 0 0 124416 0 2380 3861 26 4 57 12 0
2 1 75520 450164 688464 392212 0 7444 97280 7464 2465 3807 26 4 47 23 0
2 0 68608 417208 721452 398936 0 252 123904 252 2433 4068 27 5 56 13 0
1 1 68608 391724 743596 400604 0 7040 114176 7040 2480 3835 27 5 51 18 0
1 1 76032 404348 733084 391932 0 14976 104960 14976 2454 4065 26 5 47 22 0

Any thoughts? Could this behaviour could be by design? I'm happy to try to help troubleshoot further.

Best regards,
Chris

\
 
 \ /
  Last update: 2020-01-10 08:51    [W:0.028 / U:10.704 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site