zsmalloc: rework compaction algorithm
The zsmalloc compaction algorithm has the potential to waste some CPU
cycles, particularly when compacting pages within the same fullness group.
This is due to the way it selects the head page of the fullness list for
source and destination pages, and how it reinserts those pages during each
iteration. The algorithm may first use a page as a migration destination
and then as a migration source, leading to an unnecessary back-and-forth
movement of objects.
Consider the following fullness list:
PageA PageB PageC PageD PageE
During the first iteration, the compaction algorithm will select PageA as
the source and PageB as the destination. All of PageA's objects will be
moved to PageB, and then PageA will be released while PageB is reinserted
into the fullness list.
PageB PageC PageD PageE
During the next iteration, the compaction algorithm will again select the
head of the list as the source and destination, meaning that PageB will
now serve as the source and PageC as the destination. This will result in
the objects being moved away from PageB, the same objects that were just
moved to PageB in the previous iteration.
To prevent this avalanche effect, the compaction algorithm should not
reinsert the destination page between iterations. By doing so, the most
optimal page will continue to be used and its usage ratio will increase,
reducing internal fragmentation. The destination page should only be
reinserted into the fullness list if:
- It becomes full
- No source page is available.
TEST
====
It's very challenging to reliably test this series. I ended up developing
my own synthetic test that has 100% reproducibility. The test generates
significan fragmentation (for each size class) and then performs
compaction for each class individually and tracks the number of memcpy()
in zs_object_copy(), so that we can compare the amount work compaction
does on per-class basis.
Total amount of work (zram mm_stat objs_moved)
----------------------------------------------
Old fullness grouping, old compaction algorithm:
323977 memcpy() in zs_object_copy().
Old fullness grouping, new compaction algorithm:
262944 memcpy() in zs_object_copy().
New fullness grouping, new compaction algorithm:
213978 memcpy() in zs_object_copy().
Per-class compaction memcpy() comparison (T-test)
-------------------------------------------------
x Old fullness grouping, old compaction algorithm
+ Old fullness grouping, new compaction algorithm
N Min Max Median Avg Stddev
x 140 349 3513 2461 2314.1214 806.03271
+ 140 289 2778 2006 1878.1714 641.02073
Difference at 95.0% confidence
-435.95 +/- 170.595
-18.8387% +/- 7.37193%
(Student's t, pooled s = 728.216)
x Old fullness grouping, old compaction algorithm
+ New fullness grouping, new compaction algorithm
N Min Max Median Avg Stddev
x 140 349 3513 2461 2314.1214 806.03271
+ 140 226 2279 1644 1528.4143 524.85268
Difference at 95.0% confidence
-785.707 +/- 159.331
-33.9527% +/- 6.88516%
(Student's t, pooled s = 680.132)
Link: https://lkml.kernel.org/r/20230304034835.2082479-4-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
1 file changed
