Testing from the XFS folk revealed that there is still too much I/O from
the end of the LRU in kswapd. Previously it was considered acceptable by
VM people for a small number of pages to be written back from reclaim with
testing generally showing about 0.3% of pages reclaimed were written back
(higher if memory was low). That writing back a small number of pages is
ok has been heavily disputed for quite some time and Dave Chinner
explained it well;
It doesn't have to be a very high number to be a problem. IO
is orders of magnitude slower than the CPU time it takes to
flush a page, so the cost of making a bad flush decision is
very high. And single page writeback from the LRU is almost
always a bad flush decision.
To complicate matters, filesystems respond very differently to requests
from reclaim according to Christoph Hellwig;
xfs tries to write it back if the requester is kswapd
ext4 ignores the request if it's a delayed allocation
btrfs ignores the request
As a result, each filesystem has different performance characteristics
when under memory pressure and there are many pages being dirtied. In
some cases, the request is ignored entirely so the VM cannot depend on the
IO being dispatched.
The objective of this series is to reduce writing of filesystem-backed
pages from reclaim, play nicely with writeback that is already in progress
and throttle reclaim appropriately when writeback pages are encountered.
The assumption is that the flushers will always write pages faster than if
reclaim issues the IO.
A secondary goal is to avoid the problem whereby direct reclaim splices
two potentially deep call stacks together.
There is a potential new problem as reclaim has less control over how long
before a page in a particularly zone or container is cleaned and direct
reclaimers depend on kswapd or flusher threads to do the necessary work.
However, as filesystems sometimes ignore direct reclaim requests already,
it is not expected to be a serious issue.
Patch 1 disables writeback of filesystem pages from direct reclaim
entirely. Anonymous pages are still written.
Patch 2 removes dead code in lumpy reclaim as it is no longer able
to synchronously write pages. This hurts lumpy reclaim but
there is an expectation that compaction is used for hugepage
allocations these days and lumpy reclaim's days are numbered.
Patches 3-4 add warnings to XFS and ext4 if called from
direct reclaim. With patch 1, this "never happens" and is
intended to catch regressions in this logic in the future.
Patch 5 disables writeback of filesystem pages from kswapd unless
the priority is raised to the point where kswapd is considered
to be in trouble.
Patch 6 throttles reclaimers if too many dirty pages are being
encountered and the zones or backing devices are congested.
Patch 7 invalidates dirty pages found at the end of the LRU so they
are reclaimed quickly after being written back rather than
waiting for a reclaimer to find them
I consider this series to be orthogonal to the writeback work but it is
worth noting that the writeback work affects the viability of patch 8 in
particular.
I tested this on ext4 and xfs using fs_mark, a simple writeback test based
on dd and a micro benchmark that does a streaming write to a large mapping
(exercises use-once LRU logic) followed by streaming writes to a mix of
anonymous and file-backed mappings. The command line for fs_mark when
botted with 512M looked something like
./fs_mark -d /tmp/fsmark-2676 -D 100 -N 150 -n 150 -L 25 -t 1 -S0 -s 10485760
The number of files was adjusted depending on the amount of available
memory so that the files created was about 3xRAM. For multiple threads,
the -d switch is specified multiple times.
The test machine is x86-64 with an older generation of AMD processor with
4 cores. The underlying storage was 4 disks configured as RAID-0 as this
was the best configuration of storage I had available. Swap is on a
separate disk. Dirty ratio was tuned to 40% instead of the default of
20%.
Testing was run with and without monitors to both verify that the patches
were operating as expected and that any performance gain was real and not
due to interference from monitors.
Here is a summary of results based on testing XFS.
512M1P-xfs Files/s mean 32.69 ( 0.00%) 34.44 ( 5.08%)
512M1P-xfs Elapsed Time fsmark 51.41 48.29
512M1P-xfs Elapsed Time simple-wb 114.09 108.61
512M1P-xfs Elapsed Time mmap-strm 113.46 109.34
512M1P-xfs Kswapd efficiency fsmark 62% 63%
512M1P-xfs Kswapd efficiency simple-wb 56% 61%
512M1P-xfs Kswapd efficiency mmap-strm 44% 42%
512M-xfs Files/s mean 30.78 ( 0.00%) 35.94 (14.36%)
512M-xfs Elapsed Time fsmark 56.08 48.90
512M-xfs Elapsed Time simple-wb 112.22 98.13
512M-xfs Elapsed Time mmap-strm 219.15 196.67
512M-xfs Kswapd efficiency fsmark 54% 56%
512M-xfs Kswapd efficiency simple-wb 54% 55%
512M-xfs Kswapd efficiency mmap-strm 45% 44%
512M-4X-xfs Files/s mean 30.31 ( 0.00%) 33.33 ( 9.06%)
512M-4X-xfs Elapsed Time fsmark 63.26 55.88
512M-4X-xfs Elapsed Time simple-wb 100.90 90.25
512M-4X-xfs Elapsed Time mmap-strm 261.73 255.38
512M-4X-xfs Kswapd efficiency fsmark 49% 50%
512M-4X-xfs Kswapd efficiency simple-wb 54% 56%
512M-4X-xfs Kswapd efficiency mmap-strm 37% 36%
512M-16X-xfs Files/s mean 60.89 ( 0.00%) 65.22 ( 6.64%)
512M-16X-xfs Elapsed Time fsmark 67.47 58.25
512M-16X-xfs Elapsed Time simple-wb 103.22 90.89
512M-16X-xfs Elapsed Time mmap-strm 237.09 198.82
512M-16X-xfs Kswapd efficiency fsmark 45% 46%
512M-16X-xfs Kswapd efficiency simple-wb 53% 55%
512M-16X-xfs Kswapd efficiency mmap-strm 33% 33%
Up until 512-4X, the FSmark improvements were statistically significant.
For the 4X and 16X tests the results were within standard deviations but
just barely. The time to completion for all tests is improved which is an
important result. In general, kswapd efficiency is not affected by
skipping dirty pages.
1024M1P-xfs Files/s mean 39.09 ( 0.00%) 41.15 ( 5.01%)
1024M1P-xfs Elapsed Time fsmark 84.14 80.41
1024M1P-xfs Elapsed Time simple-wb 210.77 184.78
1024M1P-xfs Elapsed Time mmap-strm 162.00 160.34
1024M1P-xfs Kswapd efficiency fsmark 69% 75%
1024M1P-xfs Kswapd efficiency simple-wb 71% 77%
1024M1P-xfs Kswapd efficiency mmap-strm 43% 44%
1024M-xfs Files/s mean 35.45 ( 0.00%) 37.00 ( 4.19%)
1024M-xfs Elapsed Time fsmark 94.59 91.00
1024M-xfs Elapsed Time simple-wb 229.84 195.08
1024M-xfs Elapsed Time mmap-strm 405.38 440.29
1024M-xfs Kswapd efficiency fsmark 79% 71%
1024M-xfs Kswapd efficiency simple-wb 74% 74%
1024M-xfs Kswapd efficiency mmap-strm 39% 42%
1024M-4X-xfs Files/s mean 32.63 ( 0.00%) 35.05 ( 6.90%)
1024M-4X-xfs Elapsed Time fsmark 103.33 97.74
1024M-4X-xfs Elapsed Time simple-wb 204.48 178.57
1024M-4X-xfs Elapsed Time mmap-strm 528.38 511.88
1024M-4X-xfs Kswapd efficiency fsmark 81% 70%
1024M-4X-xfs Kswapd efficiency simple-wb 73% 72%
1024M-4X-xfs Kswapd efficiency mmap-strm 39% 38%
1024M-16X-xfs Files/s mean 42.65 ( 0.00%) 42.97 ( 0.74%)
1024M-16X-xfs Elapsed Time fsmark 103.11 99.11
1024M-16X-xfs Elapsed Time simple-wb 200.83 178.24
1024M-16X-xfs Elapsed Time mmap-strm 397.35 459.82
1024M-16X-xfs Kswapd efficiency fsmark 84% 69%
1024M-16X-xfs Kswapd efficiency simple-wb 74% 73%
1024M-16X-xfs Kswapd efficiency mmap-strm 39% 40%
All FSMark tests up to 16X had statistically significant improvements.
For the most part, tests are completing faster with the exception of the
streaming writes to a mixture of anonymous and file-backed mappings which
were slower in two cases
In the cases where the mmap-strm tests were slower, there was more
swapping due to dirty pages being skipped. The number of additional pages
swapped is almost identical to the fewer number of pages written from
reclaim. In other words, roughly the same number of pages were reclaimed
but swapping was slower. As the test is a bit unrealistic and stresses
memory heavily, the small shift is acceptable.
4608M1P-xfs Files/s mean 29.75 ( 0.00%) 30.96 ( 3.91%)
4608M1P-xfs Elapsed Time fsmark 512.01 492.15
4608M1P-xfs Elapsed Time simple-wb 618.18 566.24
4608M1P-xfs Elapsed Time mmap-strm 488.05 465.07
4608M1P-xfs Kswapd efficiency fsmark 93% 86%
4608M1P-xfs Kswapd efficiency simple-wb 88% 84%
4608M1P-xfs Kswapd efficiency mmap-strm 46% 45%
4608M-xfs Files/s mean 27.60 ( 0.00%) 28.85 ( 4.33%)
4608M-xfs Elapsed Time fsmark 555.96 532.34
4608M-xfs Elapsed Time simple-wb 659.72 571.85
4608M-xfs Elapsed Time mmap-strm 1082.57 1146.38
4608M-xfs Kswapd efficiency fsmark 89% 91%
4608M-xfs Kswapd efficiency simple-wb 88% 82%
4608M-xfs Kswapd efficiency mmap-strm 48% 46%
4608M-4X-xfs Files/s mean 26.00 ( 0.00%) 27.47 ( 5.35%)
4608M-4X-xfs Elapsed Time fsmark 592.91 564.00
4608M-4X-xfs Elapsed Time simple-wb 616.65 575.07
4608M-4X-xfs Elapsed Time mmap-strm 1773.02 1631.53
4608M-4X-xfs Kswapd efficiency fsmark 90% 94%
4608M-4X-xfs Kswapd efficiency simple-wb 87% 82%
4608M-4X-xfs Kswapd efficiency mmap-strm 43% 43%
4608M-16X-xfs Files/s mean 26.07 ( 0.00%) 26.42 ( 1.32%)
4608M-16X-xfs Elapsed Time fsmark 602.69 585.78
4608M-16X-xfs Elapsed Time simple-wb 606.60 573.81
4608M-16X-xfs Elapsed Time mmap-strm 1549.75 1441.86
4608M-16X-xfs Kswapd efficiency fsmark 98% 98%
4608M-16X-xfs Kswapd efficiency simple-wb 88% 82%
4608M-16X-xfs Kswapd efficiency mmap-strm 44% 42%
Unlike the other tests, the fsmark results are not statistically
significant but the min and max times are both improved and for the most
part, tests completed faster.
There are other indications that this is an improvement as well. For
example, in the vast majority of cases, there were fewer pages scanned by
direct reclaim implying in many cases that stalls due to direct reclaim
are reduced. KSwapd is scanning more due to skipping dirty pages which is
unfortunate but the CPU usage is still acceptable
In an earlier set of tests, I used blktrace and in almost all cases
throughput throughout the entire test was higher. However, I ended up
discarding those results as recording blktrace data was too heavy for my
liking.
On a laptop, I plugged in a USB stick and ran a similar tests of tests
using it as backing storage. A desktop environment was running and for
the entire duration of the tests, firefox and gnome terminal were
launching and exiting to vaguely simulate a user.
1024M-xfs Files/s mean 0.41 ( 0.00%) 0.44 ( 6.82%)
1024M-xfs Elapsed Time fsmark 2053.52 1641.03
1024M-xfs Elapsed Time simple-wb 1229.53 768.05
1024M-xfs Elapsed Time mmap-strm 4126.44 4597.03
1024M-xfs Kswapd efficiency fsmark 84% 85%
1024M-xfs Kswapd efficiency simple-wb 92% 81%
1024M-xfs Kswapd efficiency mmap-strm 60% 51%
1024M-xfs Avg wait ms fsmark 5404.53 4473.87
1024M-xfs Avg wait ms simple-wb 2541.35 1453.54
1024M-xfs Avg wait ms mmap-strm 3400.25 3852.53
The mmap-strm results were hurt because firefox launching had a tendency
to push the test out of memory. On the postive side, firefox launched
marginally faster with the patches applied. Time to completion for many
tests was faster but more importantly - the "Avg wait" time as measured by
iostat was far lower implying the system would be more responsive. It was
also the case that "Avg wait ms" on the root filesystem was lower. I
tested it manually and while the system felt slightly more responsive
while copying data to a USB stick, it was marginal enough that it could be
my imagination.
This patch: do not writeback filesystem pages in direct reclaim.
When kswapd is failing to keep zones above the min watermark, a process
will enter direct reclaim in the same manner kswapd does. If a dirty page
is encountered during the scan, this page is written to backing storage
using mapping->writepage.
This causes two problems. First, it can result in very deep call stacks,
particularly if the target storage or filesystem are complex. Some
filesystems ignore write requests from direct reclaim as a result. The
second is that a single-page flush is inefficient in terms of IO. While
there is an expectation that the elevator will merge requests, this does
not always happen. Quoting Christoph Hellwig;
The elevator has a relatively small window it can operate on,
and can never fix up a bad large scale writeback pattern.
This patch prevents direct reclaim writing back filesystem pages by
checking if current is kswapd. Anonymous pages are still written to swap
as there is not the equivalent of a flusher thread for anonymous pages.
If the dirty pages cannot be written back, they are placed back on the LRU
lists. There is now a direct dependency on dirty page balancing to
prevent too many pages in the system being dirtied which would prevent
reclaim making forward progress.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Alex Elder <aelder@sgi.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>