Hi All,
This patch-series enables zswap_store() to accept and store large
folios. The most significant contribution in this series is from the
earlier RFC submitted by Ryan Roberts [1]. Ryan's original RFC has been
migrated to mm-unstable as of 9-27-2024 in patch 6 of this series, and
adapted based on code review comments received for v7 of the current
patch-series.
[1]: [RFC PATCH v1] mm: zswap: Store large folios without splitting
https://lore.kernel.org/linux-mm/20231019110543.3284654-1-ryan.roberts@xxxxxxx/T/#u
The first few patches do the prep work for supporting large folios in
zswap_store. Patch 6 provides the main functionality to swap-out large
folios in zswap. Patch 7 adds sysfs per-order hugepages "zswpout" counters
that get incremented upon successful zswap_store of large folios:
/sys/kernel/mm/transparent_hugepage/hugepages-*kB/stats/zswpout
Patch 8 updates the documentation for the new sysfs "zswpout" counters.
This patch-series is a pre-requisite for zswap compress batching of large
folio swap-out and decompress batching of swap-ins based on
swapin_readahead(), using Intel IAA hardware acceleration, which we would
like to submit in subsequent patch-series, with performance improvement
data.
Thanks to Ying Huang for pre-posting review feedback and suggestions!
Thanks also to Nhat, Yosry, Johannes, Barry, Chengming, Usama and Ying for
their helpful feedback, data reviews and suggestions!
Co-development signoff request:
===============================
I would like to thank Ryan Roberts for his original RFC [1] and request
his co-developer signoff on patch 6 in this series. Thanks Ryan!
System setup for testing:
=========================
Testing of this patch-series was done with mm-unstable as of 9-27-2024,
commit de2fbaa6d9c3576ec7133ed02a370ec9376bf000. Data was gathered
without/with this patch-series, on an Intel Sapphire Rapids server,
dual-socket 56 cores per socket, 4 IAA devices per socket, 503 GiB RAM and
525G SSD disk partition swap. Core frequency was fixed at 2500MHz.
The vm-scalability "usemem" test was run in a cgroup whose memory.high
was fixed at 150G. The is no swap limit set for the cgroup. 30 usemem
processes were run, each allocating and writing 10G of memory, and sleeping
for 10 sec before exiting:
usemem --init-time -w -O -s 10 -n 30 10g
Other kernel configuration parameters:
zswap compressors : zstd, deflate-iaa
zswap allocator : zsmalloc
vm.page-cluster : 2
In the experiments where "deflate-iaa" is used as the zswap compressor,
IAA "compression verification" is enabled by default
(cat /sys/bus/dsa/drivers/crypto/verify_compress). Hence each IAA
compression will be decompressed internally by the "iaa_crypto" driver, the
crc-s returned by the hardware will be compared and errors reported in case
of mismatches. Thus "deflate-iaa" helps ensure better data integrity as
compared to the software compressors, and the experimental data listed
below is with verify_compress set to "1".
Total and average throughput are derived from the individual 30 processes'
throughputs reported by usemem. elapsed/sys times are measured with perf.
The vm stats and sysfs hugepages stats included with the performance data
provide details on the swapout activity to zswap/swap device.
Testing labels used in data summaries:
======================================
The data refers to these test configurations and the before/after
comparisons that they do:
before-case1:
-------------
mm-unstable 9-27-2024, CONFIG_THP_SWAP=N (compares zswap 4K vs. zswap 64K)
In this scenario, CONFIG_THP_SWAP=N results in 64K/2M folios to be split
into 4K folios that get processed by zswap.
before-case2:
-------------
mm-unstable 9-27-2024, CONFIG_THP_SWAP=Y (compares SSD swap large folios vs. zswap large folios)
In this scenario, CONFIG_THP_SWAP=Y results in zswap rejecting large
folios, which will then be stored by the SSD swap device.
after:
------
v8 of this patch-series, CONFIG_THP_SWAP=Y
The "after" is CONFIG_THP_SWAP=Y and v8 of this patch-series, that results
in 64K/2M folios to not be split, and to be processed by zswap_store.
Regression Testing:
===================
I ran vm-scalability usemem without large folios, i.e., only 4K folios with
mm-unstable and this patch-series. The main goal was to make sure that
there is no functional or performance regression wrt the earlier zswap
behavior for 4K folios, now that 4K folios will be processed by the new
zswap_store() code.
The data indicates there is no significant regression.
-------------------------------------------------------------------------------
4K folios:
==========
zswap compressor zstd zstd zstd zstd v8 zstd v8
before-case1 before-case2 after vs. vs.
case1 case2
-------------------------------------------------------------------------------
Total throughput (KB/s) 4,793,363 4,880,978 4,813,151 0% -1%
Average throughput (KB/s) 159,778 162,699 160,438 0% -1%
elapsed time (sec) 130.14 123.17 127.21 2% -3%
sys time (sec) 3,135.53 2,985.64 3,110.53 1% -4%
memcg_high 446,826 444,626 448,231
memcg_swap_fail 0 0 0
pswpout 0 0 0
pswpin 0 0 0
zswpout 48,932,107 48,931,971 48,931,584
zswpin 383 386 388
thp_swpout 0 0 0
thp_swpout_fallback 0 0 0
64kB-mthp_swpout_fallback 0 0 0
pgmajfault 3,063 3,077 3,082
swap_ra 93 94 93
swap_ra_hit 47 47 47
ZSWPOUT-64kB n/a n/a 0
SWPOUT-64kB 0 0 0
-------------------------------------------------------------------------------
Performance Testing:
====================
We list the data for 64K folios with before/after data per-compressor,
followed by the same for 2M pmd-mappable folios.
-------------------------------------------------------------------------------
64K folios: zstd:
=================
zswap compressor zstd zstd zstd zstd v8
before-case1 before-case2 after vs. vs.
case1 case2
-------------------------------------------------------------------------------
Total throughput (KB/s) 5,222,213 1,076,611 6,227,367 19% 478%
Average throughput (KB/s) 174,073 35,887 207,578 19% 478%
elapsed time (sec) 120.50 347.16 109.21 9% 69%
sys time (sec) 2,930.33 248.16 2,609.22 11% -951%
memcg_high 416,773 552,200 482,703
memcg_swap_fail 3,192,906 1,293 944
pswpout 0 40,778,448 0
pswpin 0 16 0
zswpout 48,931,583 20,903 48,931,271
zswpin 384 363 392
thp_swpout 0 0 0
thp_swpout_fallback 0 0 0
64kB-mthp_swpout_fallback 3,192,906 1,293 944
pgmajfault 3,452 3,072 3,095
swap_ra 90 87 100
swap_ra_hit 42 43 56
ZSWPOUT-64kB n/a n/a 3,057,260
SWPOUT-64kB 0 2,548,653 0
-------------------------------------------------------------------------------
-------------------------------------------------------------------------------
64K folios: deflate-iaa:
========================
zswap compressor deflate-iaa deflate-iaa deflate-iaa deflate-iaa v8
before-case1 before-case2 after vs. vs.
case1 case2
-------------------------------------------------------------------------------
Total throughput (KB/s) 5,652,608 1,089,180 6,315,000 12% 480%
Average throughput (KB/s) 188,420 36,306 210,500 12% 480%
elapsed time (sec) 102.90 343.35 91.11 11% 73%
sys time (sec) 2,246.86 213.53 1,939.31 14% -808%
memcg_high 576,104 502,907 612,505
memcg_swap_fail 4,016,117 1,407 1,660
pswpout 0 40,862,080 0
pswpin 0 20 0
zswpout 61,163,423 22,444 57,317,607
zswpin 401 368 449
thp_swpout 0 0 0
thp_swpout_fallback 0 0 0
64kB-mthp_swpout_fallback 4,016,117 1,407 1,660
pgmajfault 3,063 3,153 3,167
swap_ra 96 93 149
swap_ra_hit 46 45 89
ZSWPOUT-64kB n/a n/a 3,580,673
SWPOUT-64kB 0 2,553,880 0
-------------------------------------------------------------------------------
-------------------------------------------------------------------------------
2M folios: zstd:
================
zswap compressor zstd zstd zstd zstd v8
before-case1 before-case2 after vs. vs.
case1 case2
-------------------------------------------------------------------------------
Total throughput (KB/s) 5,895,500 1,109,694 6,460,111 10% 482%
Average throughput (KB/s) 196,516 36,989 215,337 10% 482%
elapsed time (sec) 108.77 334.28 105.92 3% 68%
sys time (sec) 2,657.14 94.88 2,436.24 8% -2468%
memcg_high 64,200 66,316 60,300
memcg_swap_fail 101,182 70 30
pswpout 0 40,166,400 0
pswpin 0 0 0
zswpout 48,931,499 36,507 48,869,236
zswpin 380 379 397
thp_swpout 0 78,450 0
thp_swpout_fallback 101,182 70 30
2MB-mthp_swpout_fallback 0 0 0
pgmajfault 3,067 3,417 4,765
swap_ra 91 90 5,073
swap_ra_hit 45 45 5,024
ZSWPOUT-2MB n/a n/a 95,408
SWPOUT-2MB 0 78,450 0
-------------------------------------------------------------------------------
-------------------------------------------------------------------------------
2M folios: deflate-iaa:
=======================
zswap compressor deflate-iaa deflate-iaa deflate-iaa deflate-iaa v8
before-case1 before-case2 after vs. vs.
case1 case2
-------------------------------------------------------------------------------
Total throughput (KB/s) 6,286,587 1,126,785 7,569,560 20% 572%
Average throughput (KB/s) 209,552 37,559 252,318 20% 572%
elapsed time (sec) 96.19 333.03 81.96 15% 75%
sys time (sec) 2,141.44 99.96 1,768.41 17% -1669%
memcg_high 99,253 64,666 75,139
memcg_swap_fail 129,074 53 73
pswpout 0 40,048,128 0
pswpin 0 0 0
zswpout 61,312,794 28,321 57,083,119
zswpin 383 406 447
thp_swpout 0 78,219 0
thp_swpout_fallback 129,074 53 73
2MB-mthp_swpout_fallback 0 0 0
pgmajfault 3,430 3,077 7,133
swap_ra 91 103 11,978
swap_ra_hit 47 46 11,920
ZSWPOUT-2MB n/a n/a 111,390
SWPOUT-2MB 0 78,219 0
-------------------------------------------------------------------------------
And finally, this is a comparison of deflate-iaa vs. zstd with v8 of this
patch-series:
---------------------------------------------
zswap_store large folios v8
Impr w/ deflate-iaa vs. zstd
64K folios 2M folios
---------------------------------------------
Throughput (KB/s) 1% 17%
elapsed time (sec) 17% 23%
sys time (sec) 26% 27%
---------------------------------------------
Conclusions based on the performance results:
=============================================
v8 wrt before-case1:
--------------------
We see significant improvements in throughput, elapsed and sys time for
zstd and deflate-iaa, when comparing before-case1 (THP_SWAP=N) vs. after
(THP_SWAP=Y) with zswap_store large folios.
v8 wrt before-case2:
--------------------
We see even more significant improvements in throughput and elapsed time
for zstd and deflate-iaa, when comparing before-case2 (large-folio-SSD)
vs. after (large-folio-zswap). The sys time increases with
large-folio-zswap as expected, due to the CPU compression time
vs. asynchronous disk write times, as pointed out by Ying and Yosry.
In before-case2, when zswap does not store large folios, only allocations
and cgroup charging due to 4K folio zswap stores count towards the cgroup
memory limit. However, in the after scenario, with the introduction of
zswap_store() of large folios, there is an added component of the zswap
compressed pool usage from large folio stores from potentially all 30
processes, that gets counted towards the memory limit. As a result, we see
higher swapout activity in the "after" data.
Summary:
========
The v8 data presented above shows that zswap_store of large folios
demonstrates good throughput/performance improvements compared to
conventional SSD swap of large folios with a sufficiently large 525G SSD
swap device. Hence, it seems reasonable for zswap_store to support large
folios, so that further performance improvements can be implemented.
In the experimental setup used in this patchset, we have enabled IAA
compress verification to ensure additional hardware data integrity CRC
checks not currently done by the software compressors. We see good
throughput/latency improvements with deflate-iaa vs. zstd with zswap_store
of large folios.
Some of the ideas for further reducing latency that have shown promise in
our experiments, are:
1) IAA compress/decompress batching.
2) Distributing compress jobs across all IAA devices on the socket.
The tests run for this patchset are using only 1 IAA device per core, that
avails of 2 compress engines on the device. In our experiments with IAA
batching, we distribute compress jobs from all cores to the 8 compress
engines available per socket. We further compress the pages in each folio
in parallel in the accelerator. As a result, we improve compress latency
and reclaim throughput.
In decompress batching, we use swapin_readahead to generate a prefetch
batch of 4K folios that we decompress in parallel in IAA.
------------------------------------------------------------------------------
IAA compress/decompress batching
Further improvements wrt v8 zswap_store Sequential
subpage store using "deflate-iaa":
"deflate-iaa" Batching "deflate-iaa-canned" [2] Batching
Additional Impr Additional Impr
64K folios 2M folios 64K folios 2M folios
------------------------------------------------------------------------------
Throughput (KB/s) 35% 34% 44% 44%
elapsed time (sec) 9% 10% 14% 17%
sys time (sec) 0.4% 4% 8% 15%
------------------------------------------------------------------------------
With zswap IAA compress/decompress batching, we are able to demonstrate
significant performance improvements and memory savings in server
scalability experiments in highly contended system scenarios under
significant memory pressure; as compared to software compressors. We hope
to submit this work in subsequent patch series. The current patch-series is
a prequisite for these future submissions.
Thanks,
Kanchana
[1] https://lore.kernel.org/linux-mm/20231019110543.3284654-1-ryan.roberts@xxxxxxx/T/#u
[2] https://patchwork.kernel.org/project/linux-crypto/cover/cover.1710969449.git.andre.glover@xxxxxxxxxxxxxxx/
Changes since v7:
=================
1) Rebased to mm-unstable as of 9-27-2024,
commit de2fbaa6d9c3576ec7133ed02a370ec9376bf000.
2) Added Nhat's 'Reviewed-by' to patches 1 and 2. Thanks Nhat!
3) Implemented one-time obj_cgroup_may_zswap and zswap_check_limits at the
start of zswap_store. Implemented one-time batch updates to cgroup zswap
charging (with total compressed bytes), zswap_stored_pages and the
memcg/vm zswpout event stats (with folio_nr_pages()) only for successful
stores at the end of zswap_store. Thanks Yosry and Johannes for guidance
on this!
4) Changed the existing zswap_pool_get() to zswap_pool_tryget(). Modified
zswap_pool_current_get() and zswap_pool_find_get() to call
zswap_pool_tryget(). Furthermore, zswap_store() obtains a reference to a
valid zswap_pool upfront by calling zswap_pool_tryget(), and errors out
if the tryget fails. Added a new zswap_pool_get() that calls
"percpu_ref_get(&pool->ref)" and is called in zswap_store_page(), as
suggested by Johannes & Yosry. Thanks both!
5) Provided a new count_objcg_events() API for batch event updates.
6) Changed "zswap_stored_pages" to atomic_long_t to support adding
folio_nr_pages() to it once a large folio is stored successfully.
7) Deleted the refactoring done in v7 for the xarray updates in
zswap_store_page(); and unwinding of stored offsets in zswap_store() in
case of errors, as suggested by Johannes.
8) Deleted the CONFIG_ZSWAP_STORE_THP_DEFAULT_ON config option and
"zswap_mthp_enabled" tunable, as recommended by Yosry, Johannes and
Nhat.
9) Replaced references to "mTHP" with "large folios"; organized
before/after data per-compressor for easier visual comparisons;
incorporated Nhat's feedback in the documentation updates; moved
changelog to the end. Thanks Johannes, Yosry and Nhat!
10) Moved the usemem testing configuration to 30 processes, each allocating
10G within a 150G memory-limit constrained cgroup, maintaining the
allocated memory for 10 sec before exiting. Thanks Ying for this
suggestion!
Changes since v6:
=================
1) Rebased to mm-unstable as of 9-23-2024,
commit acfabf7e197f7a5bedf4749dac1f39551417b049.
2) Refactored into smaller commits, as suggested by Yosry and
Chengming. Thanks both!
3) Reworded the commit log for patches 5 and 6 as per Yosry's
suggestion. Thanks Yosry!
4) Gathered data on a Sapphire Rapids server that has 823GiB SSD swap disk
partition. Also, all experiments are run with usemem --sleep 10, so that
the memory allocated by the 70 processes remains in memory
longer. Posted elapsed and sys times. Thanks to Yosry, Nhat and Ying for
their help with refining the performance characterization methodology.
5) Updated Documentation/admin-guide/mm/transhuge.rst as suggested by
Nhat. Thanks Nhat!
Changes since v5:
=================
1) Rebased to mm-unstable as of 8/29/2024,
commit 9287e4adbc6ab8fa04d25eb82e097fed877a4642.
2) Added CONFIG_ZSWAP_STORE_THP_DEFAULT_ON (off by default) to
enable/disable zswap_store() of mTHP folios. Thanks Nhat for the
suggestion to add a knob by which users can enable/disable this
change. Nhat, I hope this is along the lines of what you were
thinking.
3) Added vm-scalability usemem data with 4K folios with
CONFIG_ZSWAP_STORE_THP_DEFAULT_ON off, that I gathered to make sure
there is no regression with this change.
4) Added data with usemem with 64K and 2M THP for an alternate view of
before/after, as suggested by Yosry, so we can understand the impact
of when mTHPs are split into 4K folios in shrink_folio_list()
(CONFIG_THP_SWAP off) vs. not split (CONFIG_THP_SWAP on) and stored
in zswap. Thanks Yosry for this suggestion.
Changes since v4:
=================
1) Published before/after data with zstd, as suggested by Nhat (Thanks
Nhat for the data reviews!).
2) Rebased to mm-unstable from 8/27/2024,
commit b659edec079c90012cf8d05624e312d1062b8b87.
3) Incorporated the change in memcontrol.h that defines obj_cgroup_get() if
CONFIG_MEMCG is not defined, to resolve build errors reported by kernel
robot; as per Nhat's and Michal's suggestion to not require a separate
patch to fix the build errors (thanks both!).
4) Deleted all same-filled folio processing in zswap_store() of mTHP, as
suggested by Yosry (Thanks Yosry!).
5) Squashed the commits that define new mthp zswpout stat counters, and
invoke count_mthp_stat() after successful zswap_store()s; into a single
commit. Thanks Yosry for this suggestion!
Changes since v3:
=================
1) Rebased to mm-unstable commit 8c0b4f7b65fd1ca7af01267f491e815a40d77444.
Thanks to Barry for suggesting aligning with Ryan Roberts' latest
changes to count_mthp_stat() so that it's always defined, even when THP
is disabled. Barry, I have also made one other change in page_io.c
where count_mthp_stat() is called by count_swpout_vm_event(). I would
appreciate it if you can review this. Thanks!
Hopefully this should resolve the kernel robot build errors.
Changes since v2:
=================
1) Gathered usemem data using SSD as the backing swap device for zswap,
as suggested by Ying Huang. Ying, I would appreciate it if you can
review the latest data. Thanks!
2) Generated the base commit info in the patches to attempt to address
the kernel test robot build errors.
3) No code changes to the individual patches themselves.
Changes since RFC v1:
=====================
1) Use sysfs for zswpout mTHP stats, as per Barry Song's suggestion.
Thanks Barry!
2) Addressed some of the code review comments that Nhat Pham provided in
Ryan's initial RFC [1]:
- Added a comment about the cgroup zswap limit checks occuring once per
folio at the beginning of zswap_store().
Nhat, Ryan, please do let me know if the comments convey the summary
from the RFC discussion. Thanks!
- Posted data on running the cgroup suite's zswap kselftest.
3) Rebased to v6.11-rc3.
4) Gathered performance data with usemem and the rebased patch-series.
Kanchana P Sridhar (8):
mm: Define obj_cgroup_get() if CONFIG_MEMCG is not defined.
mm: zswap: Modify zswap_compress() to accept a page instead of a
folio.
mm: zswap: Rename zswap_pool_get() to zswap_pool_tryget().
mm: Provide a new count_objcg_events() API for batch event updates.
mm: zswap: Modify zswap_stored_pages to be atomic_long_t.
mm: zswap: Support large folios in zswap_store().
mm: swap: Count successful large folio zswap stores in hugepage
zswpout stats.
mm: Document the newly added sysfs large folios zswpout stats.
Documentation/admin-guide/mm/transhuge.rst | 8 +-
fs/proc/meminfo.c | 2 +-
include/linux/huge_mm.h | 1 +
include/linux/memcontrol.h | 24 ++
include/linux/zswap.h | 2 +-
mm/huge_memory.c | 3 +
mm/page_io.c | 1 +
mm/zswap.c | 254 +++++++++++++++------
8 files changed, 219 insertions(+), 76 deletions(-)
base-commit: de2fbaa6d9c3576ec7133ed02a370ec9376bf000
--
2.27.0
