Hi Yosry,
> -----Original Message-----
> From: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> Sent: Tuesday, October 22, 2024 5:57 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@xxxxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx;
> hannes@xxxxxxxxxxx; nphamcs@xxxxxxxxx; chengming.zhou@xxxxxxxxx;
> usamaarif642@xxxxxxxxx; ryan.roberts@xxxxxxx; Huang, Ying
> <ying.huang@xxxxxxxxx>; 21cnbao@xxxxxxxxx; akpm@xxxxxxxxxxxxxxxxxxxx;
> linux-crypto@xxxxxxxxxxxxxxx; herbert@xxxxxxxxxxxxxxxxxxx;
> davem@xxxxxxxxxxxxx; clabbe@xxxxxxxxxxxx; ardb@xxxxxxxxxx;
> ebiggers@xxxxxxxxxx; surenb@xxxxxxxxxx; Accardi, Kristen C
> <kristen.c.accardi@xxxxxxxxx>; zanussi@xxxxxxxxxx; viro@xxxxxxxxxxxxxxxxxx;
> brauner@xxxxxxxxxx; jack@xxxxxxx; mcgrof@xxxxxxxxxx; kees@xxxxxxxxxx;
> joel.granados@xxxxxxxxxx; bfoster@xxxxxxxxxx; willy@xxxxxxxxxxxxx; linux-
> fsdevel@xxxxxxxxxxxxxxx; Feghali, Wajdi K <wajdi.k.feghali@xxxxxxxxx>; Gopal,
> Vinodh <vinodh.gopal@xxxxxxxxx>
> Subject: Re: [RFC PATCH v1 00/13] zswap IAA compress batching
>
> On Thu, Oct 17, 2024 at 11:41 PM Kanchana P Sridhar
> <kanchana.p.sridhar@xxxxxxxxx> wrote:
> >
> >
> > IAA Compression Batching:
> > =========================
> >
> > This RFC patch-series introduces the use of the Intel Analytics Accelerator
> > (IAA) for parallel compression of pages in a folio, and for batched reclaim
> > of hybrid any-order batches of folios in shrink_folio_list().
> >
> > The patch-series is organized as follows:
> >
> > 1) iaa_crypto driver enablers for batching: Relevant patches are tagged
> > with "crypto:" in the subject:
> >
> > a) async poll crypto_acomp interface without interrupts.
> > b) crypto testmgr acomp poll support.
> > c) Modifying the default sync_mode to "async" and disabling
> > verify_compress by default, to facilitate users to run IAA easily for
> > comparison with software compressors.
> > d) Changing the cpu-to-iaa mappings to more evenly balance cores to IAA
> > devices.
> > e) Addition of a "global_wq" per IAA, which can be used as a global
> > resource for the socket. If the user configures 2WQs per IAA device,
> > the driver will distribute compress jobs from all cores on the
> > socket to the "global_wqs" of all the IAA devices on that socket, in
> > a round-robin manner. This can be used to improve compression
> > throughput for workloads that see a lot of swapout activity.
> >
> > 2) Migrating zswap to use async poll in zswap_compress()/decompress().
> > 3) A centralized batch compression API that can be used by swap modules.
> > 4) IAA compress batching within large folio zswap stores.
> > 5) IAA compress batching of any-order hybrid folios in
> > shrink_folio_list(). The newly added "sysctl vm.compress-batchsize"
> > parameter can be used to configure the number of folios in [1, 32] to
> > be reclaimed using compress batching.
>
> I am still digesting this series but I have some high level questions
> that I left on some patches. My intuition though is that we should
> drop (5) from the initial proposal as it's most controversial.
> Batching reclaim of unrelated folios through zswap *might* make sense,
> but it needs a broader conversation and it needs justification on its
> own merit, without the rest of the series.
Thanks for these suggestions! Sure, I can drop (5) from the initial patch-set.
Agree also, this needs a broader discussion.
I believe the 4K folios usemem30 data in this patchset does bring across
the batching reclaim benefits to provide justification on its own merit. I added
the data on batching reclaim with kernel compilation as part of the 4K folios
experiments in the IAA decompression batching patch-series [1].
Listing it here as well. I will make sure to add this data in subsequent revs.
--------------------------------------------------------------------------
Kernel compilation in tmpfs/allmodconfig, 2G max memory:
No large folios mm-unstable-10-16-2024 shrink_folio_list()
batching of folios
--------------------------------------------------------------------------
zswap compressor zstd deflate-iaa deflate-iaa
vm.compress-batchsize n/a n/a 32
vm.page-cluster 3 3 3
--------------------------------------------------------------------------
real_sec 783.87 761.69 747.32
user_sec 15,750.07 15,716.69 15,728.39
sys_sec 6,522.32 5,725.28 5,399.44
Max_RSS_KB 1,872,640 1,870,848 1,874,432
zswpout 82,364,991 97,739,600 102,780,612
zswpin 21,303,393 27,684,166 29,016,252
pswpout 13 222 213
pswpin 12 209 202
pgmajfault 17,114,339 22,421,211 23,378,161
swap_ra 4,596,035 5,840,082 6,231,646
swap_ra_hit 2,903,249 3,682,444 3,940,420
--------------------------------------------------------------------------
The performance improvements seen does depend on compression batching in
the swap modules (zswap). The implementation in patch 12 in the compress
batching series sets up this zswap compression pipeline, that takes an array of
folios and processes them in batches of 8 pages compressed in parallel in hardware.
That being said, we do see latency improvements even with reclaim batching
combined with zswap compress batching with zstd/lzo-rle/etc. I haven't done a
lot of analysis of this, but I am guessing fewer calls from the swap layer
(swap_writepage()) into zswap could have something to do with this. If we believe
that batching can be the right thing to do even for the software compressors,
I can gather batching data with zstd for v2.
[1] https://patchwork.kernel.org/project/linux-mm/cover/20241018064805.336490-1-kanchana.p.sridhar@xxxxxxxxx/
Thanks,
Kanchana
>
> >
> > IAA compress batching can be enabled only on platforms that have IAA, by
> > setting this config variable:
> >
> > CONFIG_ZSWAP_STORE_BATCHING_ENABLED="y"
> >
> > The performance testing data with usemem 30 instances shows throughput
> > gains of up to 40%, elapsed time reduction of up to 22% and sys time
> > reduction of up to 30% with IAA compression batching.
> >
> > 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.
> >
> >
> > System setup for testing:
> > =========================
> > Testing of this patch-series was done with mm-unstable as of 10-16-2024,
> > commit 817952b8be34, 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.
> >
> > 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 compressor : deflate-iaa
> > zswap allocator : zsmalloc
> > vm.page-cluster : 2,4
> >
> > IAA "compression verification" is disabled and the async poll acomp
> > interface is used in the iaa_crypto driver (the defaults with this
> > series).
> >
> >
> > Performance testing (usemem30):
> > ===============================
> >
> > 4K folios: deflate-iaa:
> > =======================
> >
> > -------------------------------------------------------------------------------
> > mm-unstable-10-16-2024 shrink_folio_list() shrink_folio_list()
> > batching of folios batching of folios
> > -------------------------------------------------------------------------------
> > zswap compressor deflate-iaa deflate-iaa deflate-iaa
> > vm.compress-batchsize n/a 1 32
> > vm.page-cluster 2 2 2
> > -------------------------------------------------------------------------------
> > Total throughput 4,470,466 5,770,824 6,363,045
> > (KB/s)
> > Average throughput 149,015 192,360 212,101
> > (KB/s)
> > elapsed time 119.24 100.96 92.99
> > (sec)
> > sys time (sec) 2,819.29 2,168.08 1,970.79
> >
> > -------------------------------------------------------------------------------
> > memcg_high 668,185 646,357 613,421
> > memcg_swap_fail 0 0 0
> > zswpout 62,991,796 58,275,673 53,070,201
> > zswpin 431 415 396
> > pswpout 0 0 0
> > pswpin 0 0 0
> > thp_swpout 0 0 0
> > thp_swpout_fallback 0 0 0
> > pgmajfault 3,137 3,085 3,440
> > swap_ra 99 100 95
> > swap_ra_hit 42 44 45
> > -------------------------------------------------------------------------------
> >
> >
> > 16k/32/64k folios: deflate-iaa:
> > ===============================
> > All three large folio sizes 16k/32/64k were enabled to "always".
> >
> > -------------------------------------------------------------------------------
> > mm-unstable- zswap_store() + shrink_folio_list()
> > 10-16-2024 batching of batching of folios
> > pages in
> > large folios
> > -------------------------------------------------------------------------------
> > zswap compr deflate-iaa deflate-iaa deflate-iaa
> > vm.compress- n/a n/a 4 8 16
> > batchsize
> > vm.page- 2 2 2 2 2
> > cluster
> > -------------------------------------------------------------------------------
> > Total throughput 7,182,198 8,448,994 8,584,728 8,729,643
> 8,775,944
> > (KB/s)
> > Avg throughput 239,406 281,633 286,157 290,988 292,531
> > (KB/s)
> > elapsed time 85.04 77.84 77.03 75.18 74.98
> > (sec)
> > sys time (sec) 1,730.77 1,527.40 1,528.52 1,473.76 1,465.97
> >
> > -------------------------------------------------------------------------------
> > memcg_high 648,125 694,188 696,004 699,728 724,887
> > memcg_swap_fail 1,550 2,540 1,627 1,577 1,517
> > zswpout 57,606,876 56,624,450 56,125,082 55,999,42 57,352,204
> > zswpin 421 406 422 400 437
> > pswpout 0 0 0 0 0
> > pswpin 0 0 0 0 0
> > thp_swpout 0 0 0 0 0
> > thp_swpout_fallback 0 0 0 0 0
> > 16kB-mthp_swpout_ 0 0 0 0 0
> > fallback
> > 32kB-mthp_swpout_ 0 0 0 0 0
> > fallback
> > 64kB-mthp_swpout_ 1,550 2,539 1,627 1,577 1,517
> > fallback
> > pgmajfault 3,102 3,126 3,473 3,454 3,134
> > swap_ra 107 144 109 124 181
> > swap_ra_hit 51 88 45 66 107
> > ZSWPOUT-16kB 2 3 4 4 3
> > ZSWPOUT-32kB 0 2 1 1 0
> > ZSWPOUT-64kB 3,598,889 3,536,556 3,506,134 3,498,324
> 3,582,921
> > SWPOUT-16kB 0 0 0 0 0
> > SWPOUT-32kB 0 0 0 0 0
> > SWPOUT-64kB 0 0 0 0 0
> > -------------------------------------------------------------------------------
> >
> >
> > 2M folios: deflate-iaa:
> > =======================
> >
> > -------------------------------------------------------------------------------
> > mm-unstable-10-16-2024 zswap_store() batching of pages
> > in pmd-mappable folios
> > -------------------------------------------------------------------------------
> > zswap compressor deflate-iaa deflate-iaa
> > vm.compress-batchsize n/a n/a
> > vm.page-cluster 2 2
> > -------------------------------------------------------------------------------
> > Total throughput 7,444,592 8,916,349
> > (KB/s)
> > Average throughput 248,153 297,211
> > (KB/s)
> > elapsed time 86.29 73.44
> > (sec)
> > sys time (sec) 1,833.21 1,418.58
> >
> > -------------------------------------------------------------------------------
> > memcg_high 81,786 89,905
> > memcg_swap_fail 82 395
> > zswpout 58,874,092 57,721,884
> > zswpin 422 458
> > pswpout 0 0
> > pswpin 0 0
> > thp_swpout 0 0
> > thp_swpout_fallback 82 394
> > pgmajfault 14,864 21,544
> > swap_ra 34,953 53,751
> > swap_ra_hit 34,895 53,660
> > ZSWPOUT-2048kB 114,815 112,269
> > SWPOUT-2048kB 0 0
> > -------------------------------------------------------------------------------
> >
> > Since 4K folios account for ~0.4% of all zswapouts when pmd-mappable
> folios
> > are enabled for usemem30, we cannot expect much improvement from
> reclaim
> > batching.
> >
> >
> > Performance testing (Kernel compilation):
> > =========================================
> >
> > As mentioned earlier, for workloads that see a lot of swapout activity, we
> > can benefit from configuring 2 WQs per IAA device, with compress jobs from
> > all same-socket cores being distributed toothe wq.1 of all IAAs on the
> > socket, with the "global_wq" developed in this patch-series.
> >
> > Although this data includes IAA decompress batching, which will be
> > submitted as a separate RFC patch-series, I am listing it here to quantify
> > the benefit of distributing compress jobs among all IAAs. The kernel
> > compilation test with "allmodconfig" is able to quantify this well:
> >
> >
> > 4K folios: deflate-iaa: kernel compilation to quantify crypto patches
> >
> ==============================================================
> =======
> >
> >
> > ------------------------------------------------------------------------------
> > IAA shrink_folio_list() compress batching and
> > swapin_readahead() decompress batching
> >
> > 1WQ 2WQ (distribute compress jobs)
> >
> > 1 local WQ (wq.0) 1 local WQ (wq.0) +
> > per IAA 1 global WQ (wq.1) per IAA
> >
> > ------------------------------------------------------------------------------
> > zswap compressor deflate-iaa deflate-iaa
> > vm.compress-batchsize 32 32
> > vm.page-cluster 4 4
> > ------------------------------------------------------------------------------
> > real_sec 746.77 745.42
> > user_sec 15,732.66 15,738.85
> > sys_sec 5,384.14 5,247.86
> > Max_Res_Set_Size_KB 1,874,432 1,872,640
> >
> > ------------------------------------------------------------------------------
> > zswpout 101,648,460 104,882,982
> > zswpin 27,418,319 29,428,515
> > pswpout 213 22
> > pswpin 207 6
> > pgmajfault 21,896,616 23,629,768
> > swap_ra 6,054,409 6,385,080
> > swap_ra_hit 3,791,628 3,985,141
> > ------------------------------------------------------------------------------
> >
> > 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.
> >
> > I would greatly appreciate code review comments for the iaa_crypto driver
> > and mm patches included in this series!
> >
> > Thanks,
> > Kanchana
> >
> >
> >
> > Kanchana P Sridhar (13):
> > crypto: acomp - Add a poll() operation to acomp_alg and acomp_req
> > crypto: iaa - Add support for irq-less crypto async interface
> > crypto: testmgr - Add crypto testmgr acomp poll support.
> > mm: zswap: zswap_compress()/decompress() can submit, then poll an
> > acomp_req.
> > crypto: iaa - Make async mode the default.
> > crypto: iaa - Disable iaa_verify_compress by default.
> > crypto: iaa - Change cpu-to-iaa mappings to evenly balance cores to
> > IAAs.
> > crypto: iaa - Distribute compress jobs to all IAA devices on a NUMA
> > node.
> > mm: zswap: Config variable to enable compress batching in
> > zswap_store().
> > mm: zswap: Create multiple reqs/buffers in crypto_acomp_ctx if
> > platform has IAA.
> > mm: swap: Add IAA batch compression API
> > swap_crypto_acomp_compress_batch().
> > mm: zswap: Compress batching with Intel IAA in zswap_store() of large
> > folios.
> > mm: vmscan, swap, zswap: Compress batching of folios in
> > shrink_folio_list().
> >
> > crypto/acompress.c | 1 +
> > crypto/testmgr.c | 70 +-
> > drivers/crypto/intel/iaa/iaa_crypto_main.c | 467 +++++++++++--
> > include/crypto/acompress.h | 18 +
> > include/crypto/internal/acompress.h | 1 +
> > include/linux/fs.h | 2 +
> > include/linux/mm.h | 8 +
> > include/linux/writeback.h | 5 +
> > include/linux/zswap.h | 106 +++
> > kernel/sysctl.c | 9 +
> > mm/Kconfig | 12 +
> > mm/page_io.c | 152 +++-
> > mm/swap.c | 15 +
> > mm/swap.h | 96 +++
> > mm/swap_state.c | 115 +++
> > mm/vmscan.c | 154 +++-
> > mm/zswap.c | 771 +++++++++++++++++++--
> > 17 files changed, 1870 insertions(+), 132 deletions(-)
> >
> >
> > base-commit: 817952b8be34aad40e07f6832fb9d1fc08961550
> > --
> > 2.27.0
> >
> >
