Hi Yosry, > -----Original Message----- > From: Yosry Ahmed <yosryahmed@xxxxxxxxxx> > Sent: Thursday, August 29, 2024 3:49 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; > Zou, Nanhai <nanhai.zou@xxxxxxxxx>; Feghali, Wajdi K > <wajdi.k.feghali@xxxxxxxxx>; Gopal, Vinodh <vinodh.gopal@xxxxxxxxx> > Subject: Re: [PATCH v6 0/3] mm: ZSWAP swap-out of mTHP folios > > On Thu, Aug 29, 2024 at 2:27 PM Kanchana P Sridhar > <kanchana.p.sridhar@xxxxxxxxx> wrote: > > > > Hi All, > > > > This patch-series enables zswap_store() to accept and store mTHP > > 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 v6.11-rc3 in patch 2/4 of this 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 > > > > Additionally, there is an attempt to modularize some of the functionality > > in zswap_store(), to make it more amenable to supporting any-order > > mTHPs. For instance, the function zswap_store_entry() stores a > zswap_entry > > in the xarray. Likewise, zswap_delete_stored_offsets() can be used to > > delete all offsets corresponding to a higher order folio stored in zswap. > > > > For accounting purposes, the patch-series adds per-order mTHP sysfs > > "zswpout" counters that get incremented upon successful zswap_store of > > an mTHP folio: > > > > /sys/kernel/mm/transparent_hugepage/hugepages-*kB/stats/zswpout > > > > A new config variable CONFIG_ZSWAP_STORE_THP_DEFAULT_ON (off by > default) > > will enable/disable zswap storing of (m)THP. When disabled, zswap will > > fallback to rejecting the mTHP folio, to be processed by the backing > > swap device. > > > > This patch-series is a precursor to ZSWAP compress batching of mTHP > > 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 RFC patch-series, with performance improvement data. > > > > Thanks to Ying Huang for pre-posting review feedback and suggestions! > > > > Thanks also to Nhat, Yosry and Barry for their helpful feedback, data > > reviews and suggestions! > > > > 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. > > > > > > Regression Testing: > > =================== > > I ran vm-scalability usemem 70 processes without mTHP, i.e., only 4K > > folios with mm-unstable and with 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, > > CONFIG_ZSWAP_STORE_THP_DEFAULT_ON is not set, and zswap_store() of > 4K > > pages goes through the newly added code path [zswap_store(), > > zswap_store_page()]. > > > > The data indicates there is no regression. > > > > ------------------------------------------------------------------------------ > > mm-unstable 8-28-2024 zswap-mTHP v6 > > CONFIG_ZSWAP_STORE_THP_DEFAULT_ON > > is not set > > ------------------------------------------------------------------------------ > > ZSWAP compressor zstd deflate- zstd deflate- > > iaa iaa > > ------------------------------------------------------------------------------ > > Throughput (KB/s) 110,775 113,010 111,550 121,937 > > sys time (sec) 1,141.72 954.87 1,131.95 828.47 > > memcg_high 140,500 153,737 139,772 134,129 > > memcg_swap_high 0 0 0 0 > > memcg_swap_fail 0 0 0 0 > > pswpin 0 0 0 0 > > pswpout 0 0 0 0 > > zswpin 675 690 682 684 > > zswpout 9,552,298 10,603,271 9,566,392 9,267,213 > > thp_swpout 0 0 0 0 > > thp_swpout_ 0 0 0 0 > > fallback > > pgmajfault 3,453 3,468 3,841 3,487 > > ZSWPOUT-64kB-mTHP n/a n/a 0 0 > > SWPOUT-64kB-mTHP 0 0 0 0 > > ------------------------------------------------------------------------------ > > > > > > Performance Testing: > > ==================== > > Testing of this patch-series was done with the v6.11-rc3 mainline, without > > and with this patch-series, on an Intel Sapphire Rapids server, > > dual-socket 56 cores per socket, 4 IAA devices per socket. > > > > The system has 503 GiB RAM, with 176GiB ZRAM (35% of available RAM) as > the > > backing swap device for ZSWAP. zstd is configured as the ZRAM compressor. > > Core frequency was fixed at 2500MHz. > > > > The vm-scalability "usemem" test was run in a cgroup whose memory.high > > was fixed at 40G. The is no swap limit set for the cgroup. Following a > > similar methodology as in Ryan Roberts' "Swap-out mTHP without splitting" > > series [2], 70 usemem processes were run, each allocating and writing 1G of > > memory: > > > > usemem --init-time -w -O -n 70 1g > > > > The vm/sysfs mTHP stats included with the performance data provide > details > > on the swapout activity to ZSWAP/swap. > > > > Other kernel configuration parameters: > > > > ZSWAP Compressors : zstd, deflate-iaa > > ZSWAP Allocator : zsmalloc > > SWAP page-cluster : 2 > > > > In the experiments where "deflate-iaa" is used as the ZSWAP compressor, > > IAA "compression verification" is enabled. 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. > > > > Throughput is derived by averaging the individual 70 processes' throughputs > > reported by usemem. sys time is measured with perf. All data points are > > averaged across 3 runs. > > > > Case 1: Baseline with CONFIG_THP_SWAP turned off, and mTHP is split in > reclaim. > > > ============================================================== > ================= > > > > In this scenario, the "before" is CONFIG_THP_SWAP set to off, that results in > > 64K/2M (m)THP to be split, and only 4K folios processed by zswap. > > > > The "after" is CONFIG_THP_SWAP set to on, and this patch-series, that > results > > in 64K/2M (m)THP to not be split, and processed by zswap. > > > > 64KB mTHP (cgroup memory.high set to 40G): > > ========================================== > > > > ------------------------------------------------------------------------------- > > v6.11-rc3 mainline zswap-mTHP Change wrt > > Baseline Baseline > > CONFIG_THP_SWAP=N CONFIG_THP_SWAP=Y > > ------------------------------------------------------------------------------- > > ZSWAP compressor zstd deflate- zstd deflate- zstd deflate- > > iaa iaa iaa > > ------------------------------------------------------------------------------- > > Throughput (KB/s) 136,113 140,044 140,363 151,938 3% 8% > > sys time (sec) 986.78 951.95 954.85 735.47 3% 23% > > memcg_high 124,183 127,513 138,651 133,884 > > memcg_swap_high 0 0 0 0 > > memcg_swap_fail 619,020 751,099 0 0 > > pswpin 0 0 0 0 > > pswpout 0 0 0 0 > > zswpin 656 569 624 639 > > zswpout 9,413,603 11,284,812 9,453,761 9,385,910 > > thp_swpout 0 0 0 0 > > thp_swpout_ 0 0 0 0 > > fallback > > pgmajfault 3,470 3,382 4,633 3,611 > > ZSWPOUT-64kB n/a n/a 590,768 586,521 > > SWPOUT-64kB 0 0 0 0 > > ------------------------------------------------------------------------------- > > > > > > 2MB PMD-THP/2048K mTHP (cgroup memory.high set to 40G): > > ======================================================= > > > > ------------------------------------------------------------------------------ > > v6.11-rc3 mainline zswap-mTHP Change wrt > > Baseline Baseline > > CONFIG_THP_SWAP=N CONFIG_THP_SWAP=Y > > ------------------------------------------------------------------------------ > > ZSWAP compressor zstd deflate- zstd deflate- zstd deflate- > > iaa iaa iaa > > ------------------------------------------------------------------------------ > > Throughput (KB/s) 164,220 172,523 165,005 174,536 0.5% 1% > > sys time (sec) 855.76 686.94 801.72 676.65 6% 1% > > memcg_high 14,628 16,247 14,951 16,096 > > memcg_swap_high 0 0 0 0 > > memcg_swap_fail 18,698 21,114 0 0 > > pswpin 0 0 0 0 > > pswpout 0 0 0 0 > > zswpin 663 665 5,333 781 > > zswpout 8,419,458 8,992,065 8,546,895 9,355,760 > > thp_swpout 0 0 0 0 > > thp_swpout_ 18,697 21,113 0 0 > > fallback > > pgmajfault 3,439 3,496 8,139 3,582 > > ZSWPOUT-2048kB n/a n/a 16,684 18,270 > > SWPOUT-2048kB 0 0 0 0 > > ----------------------------------------------------------------------------- > > > > We see improvements overall in throughput and sys time for zstd and > > deflate-iaa, when comparing before (THP_SWAP=N) vs. after > (THP_SWAP=Y). > > > > > > Case 2: Baseline with CONFIG_THP_SWAP enabled. > > ============================================== > > > > In this scenario, the "before" represents zswap rejecting mTHP, and the > mTHP > > being stored by the backing swap device. > > > > The "after" represents data with this patch-series, that results in 64K/2M > > (m)THP being processed by zswap. > > > > 64KB mTHP (cgroup memory.high set to 40G): > > ========================================== > > > > ------------------------------------------------------------------------------ > > v6.11-rc3 mainline zswap-mTHP Change wrt > > Baseline Baseline > > ------------------------------------------------------------------------------ > > ZSWAP compressor zstd deflate- zstd deflate- zstd deflate- > > iaa iaa iaa > > ------------------------------------------------------------------------------ > > Throughput (KB/s) 161,496 156,343 140,363 151,938 -13% -3% > > sys time (sec) 771.68 802.08 954.85 735.47 -24% 8% > > memcg_high 111,223 110,889 138,651 133,884 > > memcg_swap_high 0 0 0 0 > > memcg_swap_fail 0 0 0 0 > > pswpin 16 16 0 0 > > pswpout 7,471,472 7,527,963 0 0 > > zswpin 635 605 624 639 > > zswpout 1,509 1,478 9,453,761 9,385,910 > > thp_swpout 0 0 0 0 > > thp_swpout_ 0 0 0 0 > > fallback > > pgmajfault 3,616 3,430 4,633 3,611 > > ZSWPOUT-64kB n/a n/a 590,768 586,521 > > SWPOUT-64kB 466,967 470,498 0 0 > > ------------------------------------------------------------------------------ > > > > 2MB PMD-THP/2048K mTHP (cgroup memory.high set to 40G): > > ======================================================= > > > > ------------------------------------------------------------------------------ > > v6.11-rc3 mainline zswap-mTHP Change wrt > > Baseline Baseline > > ------------------------------------------------------------------------------ > > ZSWAP compressor zstd deflate- zstd deflate- zstd deflate- > > iaa iaa iaa > > ------------------------------------------------------------------------------ > > Throughput (KB/s) 192,164 194,643 165,005 174,536 -14% -10% > > sys time (sec) 823.55 830.42 801.72 676.65 3% 19% > > memcg_high 16,054 15,936 14,951 16,096 > > memcg_swap_high 0 0 0 0 > > memcg_swap_fail 0 0 0 0 > > pswpin 0 0 0 0 > > pswpout 8,629,248 8,628,907 0 0 > > zswpin 560 645 5,333 781 > > zswpout 1,416 1,503 8,546,895 9,355,760 > > thp_swpout 16,854 16,853 0 0 > > thp_swpout_ 0 0 0 0 > > fallback > > pgmajfault 3,341 3,574 8,139 3,582 > > ZSWPOUT-2048kB n/a n/a 16,684 18,270 > > SWPOUT-2048kB 16,854 16,853 0 0 > > ------------------------------------------------------------------------------ > > > > In the "Before" scenario, when zswap does not store mTHP, only allocations > > count towards the cgroup memory limit. However, in the "After" scenario, > > with the introduction of zswap_store() mTHP, both, allocations as well as > > the zswap compressed pool usage from all 70 processes are counted > towards > > the memory limit. As a result, we see higher swapout activity in the > > "After" data. Hence, more time is spent doing reclaim as the zswap cgroup > > charge leads to more frequent memory.high breaches. > > > > This causes degradation in throughput and sys time with zswap mTHP, more > so > > in case of zstd than deflate-iaa. Compress latency could play a part in > > this - when there is more swapout activity happening, a slower compressor > > would cause allocations to stall for any/all of the 70 processes. > > We are basically comparing zram with zswap in this case, and it's not > fair because, as you mentioned, the zswap compressed data is being > accounted for while the zram compressed data isn't. I am not really > sure how valuable these test results are. Even if we remove the cgroup > accounting from zswap, we won't see an improvement, we should expect a > similar performance to zram. > > I think the test results that are really valuable are case 1, where > zswap users are currently disabling CONFIG_THP_SWAP, and get to enable > it after this series. > > If we really want to compare CONFIG_THP_SWAP on before and after, it > should be with SSD because that's a more conventional setup. In this > case the users that have CONFIG_THP_SWAP=y only experience the > benefits of zswap with this series. You mentioned experimenting with > usemem to keep the memory allocated longer so that you're able to have > a fair test with the small SSD swap setup. Did that work? Thanks, these are good points. I ran this experiment with mm-unstable 9-17-2024, commit 248ba8004e76eb335d7e6079724c3ee89a011389. Data is based on average of 3 runs of the vm-scalability "usemem" test. 4G SSD backing zswap, each process sleeps before exiting ======================================================== 64KB mTHP (cgroup memory.high set to 60G, no swap limit): ========================================================= CONFIG_THP_SWAP=Y Sapphire Rapids server with 503 GiB RAM and 4G SSD swap backing device for zswap. Experiment 1: Each process sleeps for 0 sec after allocating memory (usemem --init-time -w -O --sleep 0 -n 70 1g): ------------------------------------------------------------------------------- mm-unstable 9-17-2024 zswap-mTHP v6 Change wrt Baseline Baseline "before" "after" (sleep 0) ------------------------------------------------------------------------------- ZSWAP compressor zstd deflate- zstd deflate- zstd deflate- iaa iaa iaa ------------------------------------------------------------------------------- Throughput (KB/s) 296,684 274,207 359,722 390,162 21% 42% sys time (sec) 92.67 93.33 251.06 237.56 -171% -155% memcg_high 3,503 3,769 44,425 27,154 memcg_swap_fail 0 0 115,814 141,936 pswpin 17 0 0 0 pswpout 370,853 393,232 0 0 zswpin 693 123 666 667 zswpout 1,484 123 1,366,680 1,199,645 thp_swpout 0 0 0 0 thp_swpout_ 0 0 0 0 fallback pgmajfault 3,384 2,951 3,656 3,468 ZSWPOUT-64kB n/a n/a 82,940 73,121 SWPOUT-64kB 23,178 24,577 0 0 ------------------------------------------------------------------------------- Experiment 2: Each process sleeps for 10 sec after allocating memory (usemem --init-time -w -O --sleep 10 -n 70 1g): ------------------------------------------------------------------------------- mm-unstable 9-17-2024 zswap-mTHP v6 Change wrt Baseline Baseline "before" "after" (sleep 10) ------------------------------------------------------------------------------- ZSWAP compressor zstd deflate- zstd deflate- zstd deflate- iaa iaa iaa ------------------------------------------------------------------------------- Throughput (KB/s) 86,744 93,730 157,528 113,110 82% 21% sys time (sec) 308.87 315.29 477.55 629.98 -55% -100% memcg_high 169,450 188,700 143,691 177,887 memcg_swap_fail 10,131,859 9,740,646 18,738,715 19,528,110 pswpin 17 16 0 0 pswpout 1,154,779 1,210,485 0 0 zswpin 711 659 1,016 736 zswpout 70,212 50,128 1,235,560 1,275,917 thp_swpout 0 0 0 0 thp_swpout_ 0 0 0 0 fallback pgmajfault 6,120 6,291 8,789 6,474 ZSWPOUT-64kB n/a n/a 67,587 68,912 SWPOUT-64kB 72,174 75,655 0 0 ------------------------------------------------------------------------------- Conclusions from the experiments: ================================= 1) zswap-mTHP improves throughput as compared to the baseline, for zstd and deflate-iaa. 2) Yosry's theory is proved correct in the 4G constrained swap setup. When the processes are constrained to sleep 10 sec after allocating memory, thereby keeping the memory allocated longer, the "Baseline" or "before" with mTHP getting stored in SSD shows a degradation of 71% in throughput and 238% in sys time, as compared to the "Baseline" with sleep 0 that benefits from serialization of disk IO not allowing all processes to allocate memory at the same time. 3) In the 4G SSD "sleep 0" case, zswap-mTHP shows an increase in sys time due to the cgroup charging and consequently higher memcg.high breaches and swapout activity. However, the "sleep 10" case's sys time seems to degrade less, and the memcg.high breaches and swapout activity are almost similar between the before/after (confirming Yosry's hypothesis). Further, the memcg_swap_fail activity in the "after" scenario is almost 2X that of the "before". This indicates failure to obtain swap offsets, resulting in the folio remaining active in memory. I tried to better understand this through the 64k mTHP swpout_fallback stats in the "sleep 10" zstd experiments: -------------------------------------------------------------- "before" "after" -------------------------------------------------------------- 64k mTHP swpout_fallback 627,308 897,407 64k folio swapouts 72,174 67,587 [p|z]swpout events due to 64k mTHP 1,154,779 1,081,397 4k folio swapouts 70,212 154,163 -------------------------------------------------------------- The data indicates a higher # of 64k folio swpout_fallback with zswap-mTHP, that co-relates with the higher memcg_swap_fail counts and 4k folio swapouts with zswap-mTHP. Could the root-cause be fragmentation of the swap space due to zswap swapout being faster than SSD swapout? > > I am hoping Nhat or Johannes would shed some light on whether they > usually have CONFIG_THP_SWAP enabled or not with zswap. I am trying to > figure out if any reasonable setups enable CONFIG_THP_SWAP with zswap. > Otherwise the testing results from case 1 should be sufficient. > > > > > In my opinion, even though the test set up does not provide an accurate > > way for a direct before/after comparison (because of zswap usage being > > counted in cgroup, hence towards the memory.high), it still seems > > reasonable for zswap_store to support (m)THP, so that further performance > > improvements can be implemented. > > This is only referring to the results of case 2, right? To begin with, yes. With IAA batching, we can submit say, up to 8 pages in an mTHP for parallel compression in hardware. We have also implemented batching of any-order folios (e.g. mix of 4K/16K/64K/.. folios) reclaimed in the shrink_folio_list() -- swap_writepage() path, that demonstrates performance and memory savings improvements with IAA. > > Honestly, I wouldn't want to merge mTHP swapout support on its own > just because it enables further performance improvements without > having actual patches for them. But I don't think this captures the > results accurately as it dismisses case 1 results (which I think are > more reasonable). Based on the latest set of data, we do see consistent throughput improvements with zswap mTHP swapout using zstd and deflate-IAA, as compared to a baseline where mTHP are swapped to disk (CONFIG_THP_SWP=y). The Intel IAA batching related patches would enable the additional performance improvements I was referring to, only for configurations that have the hardware acceleration, without impacting performance of software compressors. Hence, I was thinking we could separate the patch-sets as: 1) zswap-mTHP swapout that could benefit all compressors: this patch series. 2) Additional IAA batching performance improvements that would only benefit users of IAA. I would appreciate your thoughts on this. Thanks, Kanchana > > Thnaks
