IAA Compression Batching with acomp Request Chaining:
=====================================================
This patch-series introduces the use of the Intel Analytics Accelerator
(IAA) for parallel batch compression of pages in large folios to improve
zswap swapout latency, resulting in sys time reduction by 22% (usemem30)
and by 27% (kernel compilation); as well as a 30% increase in usemem30
throughput with IAA batching as compared to zstd.
The patch-series is organized as follows:
1) crypto acomp & iaa_crypto driver enablers for batching: Relevant
patches are tagged with "crypto:" in the subject:
Patch 1) Adds new acomp request chaining framework and interface based
on Herbert Xu's ahash reference implementation in "[PATCH 2/6]
crypto: hash - Add request chaining API" [1]. acomp algorithms
can use request chaining through these interfaces:
Setup the request chain:
acomp_reqchain_init()
acomp_request_chain()
Process the request chain:
acomp_do_req_chain(): synchronously (sequentially)
acomp_do_async_req_chain(): asynchronously using submit/poll
ops (in parallel)
Patch 2) Adds acomp_alg/crypto_acomp interfaces for batch_compress(),
batch_decompress() and get_batch_size(), that swap modules can
invoke using the new batching API crypto_acomp_batch_compress(),
crypto_acomp_batch_decompress() and crypto_acomp_batch_size().
Additionally, crypto acomp provides a new
acomp_has_async_batching() interface to query for these API
before allocating batching resources for a given compressor in
zswap/zram.
Patch 3) New CRYPTO_ACOMP_REQ_POLL acomp_req flag to act as a gate for
async poll mode in iaa_crypto.
Patch 4) iaa-crypto driver implementations for sync/async
crypto_acomp_batch_compress() and
crypto_acomp_batch_decompress() developed using request
chaining. If the iaa_crypto driver is set up for 'async'
sync_mode, these batching implementations deploy the
asynchronous request chaining implementation. 'async' is the
recommended mode for realizing the benefits of IAA parallelism.
If iaa_crypto is set up for 'sync' sync_mode, the synchronous
version of the request chaining API is used.
The "iaa_acomp_fixed_deflate" algorithm registers these
implementations for its "batch_compress" and "batch_decompress"
interfaces respectively and opts in with CRYPTO_ALG_REQ_CHAIN.
Further, iaa_crypto provides an implementation for the
"get_batch_size" interface: this returns the
IAA_CRYPTO_MAX_BATCH_SIZE constant that iaa_crypto defines
currently as 8U for IAA compression algorithms (iaa_crypto can
change this if needed as we optimize our batching algorithm).
Patch 5) Modifies the default iaa_crypto driver mode to async, now that
iaa_crypto provides a truly async mode that gives
significantly better latency than sync mode for the batching
use case.
Patch 6) Disables verify_compress by default, to facilitate users to
run IAA easily for comparison with software compressors.
Patch 7) Reorganizes the iaa_crypto driver code into logically related
sections and avoids forward declarations, in order to facilitate
Patch 8. This patch makes no functional changes.
Patch 8) Makes a major infrastructure change in the iaa_crypto driver,
to map IAA devices/work-queues to cores based on packages
instead of NUMA nodes. This doesn't impact performance on
the Sapphire Rapids system used for performance
testing. However, this change fixes functional problems we
found on Granite Rapids in internal validation, where the
number of NUMA nodes is greater than the number of packages,
which was resulting in over-utilization of some IAA devices
and non-usage of other IAA devices as per the current NUMA
based mapping infrastructure.
This patch also eliminates duplication of device wqs in
per-cpu wq_tables, thereby saving 140MiB on a 384 cores
Granite Rapids server with 8 IAAs. Submitting this change now
so that it can go through code reviews before it can be merged.
Patch 9) Builds upon the new infrastructure for mapping IAAs to cores
based on packages, and enables configuring a "global_wq" per
IAA, which can be used as a global resource for compress jobs
for the package. If the user configures 2WQs per IAA device,
the driver will distribute compress jobs from all cores on the
package to the "global_wqs" of all the IAA devices on that
package, in a round-robin manner. This can be used to improve
compression throughput for workloads that see a lot of swapout
activity.
2) zswap modifications to enable compress batching in zswap_store()
of large folios (including pmd-mappable folios):
Patch 10) Defines a zswap-specific ZSWAP_MAX_BATCH_SIZE (currently set
as 8U) to denote the maximum number of acomp_ctx batching
resources. Further, the "struct crypto_acomp_ctx" is modified
to contain a configurable number of acomp_reqs and buffers.
The cpu hotplug onlining code will query
acomp_has_async_batching() and if this returns "true", will
further get the compressor defined maximum batch size, and
will use the minimum of zswap's upper limit and the
compressor's maximum batch size to allocate
acomp_reqs/buffers if the acomp supports batching, and 1
acomp_req/buffer if not.
Patch 11) Restructures & simplifies zswap_store() to make it amenable
for batching. Moves the loop over the folio's pages to a new
zswap_store_folio(), which in turn allocates zswap entries
for all folio pages upfront, before proceeding to call a
newly added zswap_compress_folio(), which simply calls
zswap_compress() for each folio page.
Patch 12) Finally, this patch modifies zswap_compress_folio() to detect
if the pool's acomp_ctx has batching resources. If so, the
"acomp_ctx->nr_reqs" becomes the batch size to use to call
crypto_acomp_batch_compress() for every "acomp_ctx->nr_reqs"
pages in the large folio. The crypto API calls into the new
iaa_crypto "iaa_comp_acompress_batch()" that does batching
with request chaining. Upon successful compression of a
batch, the compressed buffers are stored in zpool.
With v5 of this patch series, the IAA compress batching feature will be
enabled seamlessly on Intel platforms that have IAA by selecting
'deflate-iaa' as the zswap compressor, and using the iaa_crypto 'async'
sync_mode driver attribute.
[1]: https://lore.kernel.org/linux-crypto/677614fbdc70b31df2e26483c8d2cd1510c8af91.1730021644.git.herbert@xxxxxxxxxxxxxxxxxxx/
System setup for testing:
=========================
Testing of this patch-series was done with mm-unstable as of 12-20-2024,
commit 5555a83c82d6, without and with this patch-series.
Data was gathered 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.
Other kernel configuration parameters:
zswap compressor : zstd, deflate-iaa
zswap allocator : zsmalloc
vm.page-cluster : 0, 2
IAA "compression verification" is disabled and IAA is run in the async
mode (the defaults with this series). 2WQs are configured per IAA
device. Compress jobs from all cores on a socket are distributed among all
4 IAA devices on the same socket.
I ran experiments with these workloads:
1) usemem 30 processes with these large folios enabled to "always":
- 16k/32k/64k
- 2048k
2) Kernel compilation allmodconfig with 2G max memory, 32 threads, run in
tmpfs with these large folios enabled to "always":
- 16k/32k/64k
IAA compress batching performance: sync vs. async request chaining:
===================================================================
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
"async polling" here refers to the v4 implementation of batch compression
without request chaining, which is used as baseline to compare the request
chaining implementations in v5.
These are the latencies measured using bcc profiling with bpftrace for the
various iaa_crypto modes:
-------------------------------------------------------------------------------
usemem30: 16k/32k/64k Folios crypto_acomp_batch_compress() latency
iaa_crypto batching count mean p50 p99
implementation (ns) (ns) (ns)
-------------------------------------------------------------------------------
async polling 5,210,702 10,083 9,675 17,488
sync request chaining 5,396,532 33,391 32,977 39,426
async request chaining 5,509,777 9,959 9,611 16,590
-------------------------------------------------------------------------------
This demonstrates that async request chaining doesn't cause IAA compress
batching performance regression wrt the v4 implementation without request
chaining.
Performance testing (usemem30):
===============================
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
16k/32/64k folios: usemem30: zstd:
==================================
-------------------------------------------------------------------------------
mm-unstable-12-20-2024 v5 of this patch-series
-------------------------------------------------------------------------------
zswap compressor zstd zstd
vm.page-cluster 2 2
-------------------------------------------------------------------------------
Total throughput (KB/s) 6,143,774 6,180,657
Avg throughput (KB/s) 204,792 206,021
elapsed time (sec) 110.45 112.02
sys time (sec) 2,628.55 2,684.53
-------------------------------------------------------------------------------
memcg_high 469,269 481,665
memcg_swap_fail 1,198 910
zswpout 48,932,319 48,931,447
zswpin 384 398
pswpout 0 0
pswpin 0 0
thp_swpout 0 0
thp_swpout_fallback 0 0
16kB-swpout_fallback 0 0
32kB_swpout_fallback 0 0
64kB_swpout_fallback 1,198 910
pgmajfault 3,459 3,090
swap_ra 96 100
swap_ra_hit 48 54
ZSWPOUT-16kB 2 2
ZSWPOUT-32kB 2 0
ZSWPOUT-64kB 3,057,060 3,057,286
SWPOUT-16kB 0 0
SWPOUT-32kB 0 0
SWPOUT-64kB 0 0
-------------------------------------------------------------------------------
16k/32/64k folios: usemem30: deflate-iaa:
=========================================
-------------------------------------------------------------------------------
mm-unstable-12-20-2024 v5 of this patch-series
-------------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vm.page-cluster 2 2 vs. vs.
Seq zstd
-------------------------------------------------------------------------------
Total throughput (KB/s) 7,679,064 8,027,314 5% 30%
Avg throughput (KB/s) 255,968 267,577 5% 30%
elapsed time (sec) 90.82 87.53 -4% -22%
sys time (sec) 2,205.73 2,099.80 -5% -22%
-------------------------------------------------------------------------------
memcg_high 716,670 722,693
memcg_swap_fail 1,187 1,251
zswpout 64,511,695 64,510,499
zswpin 483 477
pswpout 0 0
pswpin 0 0
thp_swpout 0 0
thp_swpout_fallback 0 0
16kB-swpout_fallback 0 0
32kB_swpout_fallback 0 0
64kB_swpout_fallback 1,187 1,251
pgmajfault 3,180 3,187
swap_ra 175 155
swap_ra_hit 114 76
ZSWPOUT-16kB 5 3
ZSWPOUT-32kB 1 2
ZSWPOUT-64kB 4,030,709 4,030,573
SWPOUT-16kB 0 0
SWPOUT-32kB 0 0
SWPOUT-64kB 0 0
-------------------------------------------------------------------------------
2M folios: usemem30: zstd:
==========================
-------------------------------------------------------------------------------
mm-unstable-12-20-2024 v5 of this patch-series
-------------------------------------------------------------------------------
zswap compressor zstd zstd
vm.page-cluster 2 2
-------------------------------------------------------------------------------
Total throughput (KB/s) 6,643,427 6,534,525
Avg throughput (KB/s) 221,447 217,817
elapsed time (sec) 102.92 104.44
sys time (sec) 2,332.67 2,415.00
-------------------------------------------------------------------------------
memcg_high 61,999 60,770
memcg_swap_fail 37 47
zswpout 48,934,491 48,934,952
zswpin 386 404
pswpout 0 0
pswpin 0 0
thp_swpout 0 0
thp_swpout_fallback 37 47
pgmajfault 5,010 4,646
swap_ra 5,836 4,692
swap_ra_hit 5,790 4,640
ZSWPOUT-2048kB 95,529 95,520
SWPOUT-2048kB 0 0
-------------------------------------------------------------------------------
2M folios: usemem30: deflate-iaa:
=================================
-------------------------------------------------------------------------------
mm-unstable-12-20-2024 v5 of this patch-series
-------------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vm.page-cluster 2 2 vs. vs.
Seq zstd
-------------------------------------------------------------------------------
Total throughput (KB/s) 8,197,457 8,427,981 3% 29%
Avg throughput (KB/s) 273,248 280,932 3% 29%
elapsed time (sec) 86.79 83.45 -4% -20%
sys time (sec) 2,044.02 1,925.84 -6% -20%
-------------------------------------------------------------------------------
memcg_high 94,008 88,809
memcg_swap_fail 50 57
zswpout 64,521,910 64,520,405
zswpin 421 452
pswpout 0 0
pswpin 0 0
thp_swpout 0 0
thp_swpout_fallback 50 57
pgmajfault 9,658 8,958
swap_ra 19,633 17,341
swap_ra_hit 19,579 17,278
ZSWPOUT-2048kB 125,916 125,913
SWPOUT-2048kB 0 0
-------------------------------------------------------------------------------
Performance testing (Kernel compilation, allmodconfig):
=======================================================
The experiments with kernel compilation test, 32 threads, in tmpfs use the
"allmodconfig" that takes ~12 minutes, and has considerable swapout
activity. The cgroup's memory.max is set to 2G.
16k/32k/64k folios: Kernel compilation/allmodconfig:
====================================================
w/o: mm-unstable-12-20-2024
-------------------------------------------------------------------------------
w/o v5 w/o v5
-------------------------------------------------------------------------------
zswap compressor zstd zstd deflate-iaa deflate-iaa
vm.page-cluster 0 0 0 0
-------------------------------------------------------------------------------
real_sec 792.04 793.92 783.43 766.93
user_sec 15,781.73 15,772.48 15,753.22 15,766.53
sys_sec 5,302.83 5,308.05 3,982.30 3,853.21
-------------------------------------------------------------------------------
Max_Res_Set_Size_KB 1,871,908 1,873,368 1,871,836 1,873,168
-------------------------------------------------------------------------------
memcg_high 0 0 0 0
memcg_swap_fail 0 0 0 0
zswpout 90,775,917 91,653,816 106,964,482 110,380,500
zswpin 26,099,486 26,611,908 31,598,420 32,618,221
pswpout 48 96 331 331
pswpin 48 89 320 310
thp_swpout 0 0 0 0
thp_swpout_fallback 0 0 0 0
16kB_swpout_fallback 0 0 0 0
32kB_swpout_fallback 0 0 0 0
64kB_swpout_fallback 0 2,337 7,943 5,512
pgmajfault 27,858,798 28,438,518 33,970,455 34,999,918
swap_ra 0 0 0 0
swap_ra_hit 2,173 2,913 2,192 5,248
ZSWPOUT-16kB 1,292,865 1,306,214 1,463,397 1,483,056
ZSWPOUT-32kB 695,446 705,451 830,676 829,992
ZSWPOUT-64kB 2,938,716 2,958,250 3,520,199 3,634,972
SWPOUT-16kB 0 0 0 0
SWPOUT-32kB 0 0 0 0
SWPOUT-64kB 3 6 20 19
-------------------------------------------------------------------------------
Summary:
========
The performance testing data with usemem 30 processes and kernel
compilation test show 30% throughput gains and 22% sys time reduction
(usemem30) and 27% sys time reduction (kernel compilation) with
zswap_store() large folios using IAA compress batching as compared to
zstd.
The iaa_crypto wq stats will show almost the same number of compress calls
for wq.1 of all IAA devices. wq.0 will handle decompress calls exclusively.
We see a latency reduction of 2.5% by distributing compress jobs among all
IAA devices on the socket (based on v1 data).
We can expect to see even more significant performance and throughput
improvements if we use the parallelism offered by IAA to do reclaim
batching of 4K/large folios (really any-order folios), and using the
zswap_store() high throughput compression to batch-compress pages
comprising these folios, not just batching within large folios. This is the
reclaim batching patch 13 in v1, which will be submitted in a separate
patch-series.
Our internal validation of IAA compress/decompress batching in highly
contended Sapphire Rapids server setups with workloads running on 72 cores
for ~25 minutes under stringent memory limit constraints have shown up to
50% reduction in sys time and 3.5% reduction in workload run time as
compared to software compressors.
Changes since v4:
=================
1) Rebased to mm-unstable as of 12-20-2024, commit 5555a83c82d6.
2) Added acomp request chaining, as suggested by Herbert. Thanks Herbert!
3) Implemented IAA compress batching using request chaining.
4) zswap_store() batching simplifications suggested by Chengming, Yosry and
Nhat, thanks to all!
- New zswap_compress_folio() that is called by zswap_store().
- Move the loop over folio's pages out of zswap_store() and into a
zswap_store_folio() that stores all pages.
- Allocate all zswap entries for the folio upfront.
- Added zswap_batch_compress().
- Branch to call zswap_compress() or zswap_batch_compress() inside
zswap_compress_folio().
- All iterations over pages kept in same function level.
- No helpers other than the newly added zswap_store_folio() and
zswap_compress_folio().
Changes since v3:
=================
1) Rebased to mm-unstable as of 11-18-2024, commit 5a7056135bb6.
2) Major re-write of iaa_crypto driver's mapping of IAA devices to cores,
based on packages instead of NUMA nodes.
3) Added acomp_has_async_batching() API to crypto acomp, that allows
zswap/zram to query if a crypto_acomp has registered batch_compress and
batch_decompress interfaces.
4) Clear the poll bits on the acomp_reqs passed to
iaa_comp_a[de]compress_batch() so that a module like zswap can be
confident about the acomp_reqs[0] not having the poll bit set before
calling the fully synchronous API crypto_acomp_[de]compress().
Herbert, I would appreciate it if you can review changes 2-4; in patches
1-8 in v4. I did not want to introduce too many iaa_crypto changes in
v4, given that patch 7 is already making a major change. I plan to work
on incorporating the request chaining using the ahash interface in v5
(I need to understand the basic crypto ahash better). Thanks Herbert!
5) Incorporated Johannes' suggestion to not have a sysctl to enable
compress batching.
6) Incorporated Yosry's suggestion to allocate batching resources in the
cpu hotplug onlining code, since there is no longer a sysctl to control
batching. Thanks Yosry!
7) Incorporated Johannes' suggestions related to making the overall
sequence of events between zswap_store() and zswap_batch_store() similar
as much as possible for readability and control flow, better naming of
procedures, avoiding forward declarations, not inlining error path
procedures, deleting zswap internal details from zswap.h, etc. Thanks
Johannes, really appreciate the direction!
I have tried to explain the minimal future-proofing in terms of the
zswap_batch_store() signature and the definition of "struct
zswap_batch_store_sub_batch" in the comments for this struct. I hope the
new code explains the control flow a bit better.
Changes since v2:
=================
1) Rebased to mm-unstable as of 11-5-2024, commit 7994b7ea6ac8.
2) Fixed an issue in zswap_create_acomp_ctx() with checking for NULL
returned by kmalloc_node() for acomp_ctx->buffers and for
acomp_ctx->reqs.
3) Fixed a bug in zswap_pool_can_batch() for returning true if
pool->can_batch_comp is found to be equal to BATCH_COMP_ENABLED, and if
the per-cpu acomp_batch_ctx tests true for batching resources having
been allocated on this cpu. Also, changed from per_cpu_ptr() to
raw_cpu_ptr().
4) Incorporated the zswap_store_propagate_errors() compilation warning fix
suggested by Dan Carpenter. Thanks Dan!
5) Replaced the references to SWAP_CRYPTO_SUB_BATCH_SIZE in comments in
zswap.h, with SWAP_CRYPTO_BATCH_SIZE.
Changes since v1:
=================
1) Rebased to mm-unstable as of 11-1-2024, commit 5c4cf96cd702.
2) Incorporated Herbert's suggestions to use an acomp_req flag to indicate
async/poll mode, and to encapsulate the polling functionality in the
iaa_crypto driver. Thanks Herbert!
3) Incorporated Herbert's and Yosry's suggestions to implement the batching
API in iaa_crypto and to make its use seamless from zswap's
perspective. Thanks Herbert and Yosry!
4) Incorporated Yosry's suggestion to make it more convenient for the user
to enable compress batching, while minimizing the memory footprint
cost. Thanks Yosry!
5) Incorporated Yosry's suggestion to de-couple the shrink_folio_list()
reclaim batching patch from this series, since it requires a broader
discussion.
I would greatly appreciate code review comments for the iaa_crypto driver
and mm patches included in this series!
Thanks,
Kanchana
Kanchana P Sridhar (12):
crypto: acomp - Add synchronous/asynchronous acomp request chaining.
crypto: acomp - Define new interfaces for compress/decompress
batching.
crypto: iaa - Add an acomp_req flag CRYPTO_ACOMP_REQ_POLL to enable
async mode.
crypto: iaa - Implement batch_compress(), batch_decompress() API in
iaa_crypto.
crypto: iaa - Make async mode the default.
crypto: iaa - Disable iaa_verify_compress by default.
crypto: iaa - Re-organize the iaa_crypto driver code.
crypto: iaa - Map IAA devices/wqs to cores based on packages instead
of NUMA.
crypto: iaa - Distribute compress jobs from all cores to all IAAs on a
package.
mm: zswap: Allocate pool batching resources if the crypto_alg supports
batching.
mm: zswap: Restructure & simplify zswap_store() to make it amenable
for batching.
mm: zswap: Compress batching with Intel IAA in zswap_store() of large
folios.
crypto/acompress.c | 287 ++++
drivers/crypto/intel/iaa/iaa_crypto.h | 27 +-
drivers/crypto/intel/iaa/iaa_crypto_main.c | 1697 +++++++++++++++-----
include/crypto/acompress.h | 157 ++
include/crypto/algapi.h | 10 +
include/crypto/internal/acompress.h | 29 +
include/linux/crypto.h | 31 +
mm/zswap.c | 406 +++--
8 files changed, 2103 insertions(+), 541 deletions(-)
base-commit: 5555a83c82d66729e4abaf16ae28d6bd81f9a64a
--
2.27.0
