From a4103262b01a1b8704b37c01c7c813df91b7b119 Mon Sep 17 00:00:00 2001 From: Yu Zhao Date: Sun, 18 Sep 2022 01:59:58 -0600 Subject: [PATCH 01/29] mm: x86, arm64: add arch_has_hw_pte_young() MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Patch series "Multi-Gen LRU Framework", v14. What's new ========== 1. OpenWrt, in addition to Android, Arch Linux Zen, Armbian, ChromeOS, Liquorix, post-factum and XanMod, is now shipping MGLRU on 5.15. 2. Fixed long-tailed direct reclaim latency seen on high-memory (TBs) machines. The old direct reclaim backoff, which tries to enforce a minimum fairness among all eligible memcgs, over-swapped by about (total_mem>>DEF_PRIORITY)-nr_to_reclaim. The new backoff, which pulls the plug on swapping once the target is met, trades some fairness for curtailed latency: https://lore.kernel.org/r/20220918080010.2920238-10-yuzhao@google.com/ 3. Fixed minior build warnings and conflicts. More comments and nits. TLDR ==== The current page reclaim is too expensive in terms of CPU usage and it often makes poor choices about what to evict. This patchset offers an alternative solution that is performant, versatile and straightforward. Patchset overview ================= The design and implementation overview is in patch 14: https://lore.kernel.org/r/20220918080010.2920238-15-yuzhao@google.com/ 01. mm: x86, arm64: add arch_has_hw_pte_young() 02. mm: x86: add CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG Take advantage of hardware features when trying to clear the accessed bit in many PTEs. 03. mm/vmscan.c: refactor shrink_node() 04. Revert "include/linux/mm_inline.h: fold __update_lru_size() into its sole caller" Minor refactors to improve readability for the following patches. 05. mm: multi-gen LRU: groundwork Adds the basic data structure and the functions that insert pages to and remove pages from the multi-gen LRU (MGLRU) lists. 06. mm: multi-gen LRU: minimal implementation A minimal implementation without optimizations. 07. mm: multi-gen LRU: exploit locality in rmap Exploits spatial locality to improve efficiency when using the rmap. 08. mm: multi-gen LRU: support page table walks Further exploits spatial locality by optionally scanning page tables. 09. mm: multi-gen LRU: optimize multiple memcgs Optimizes the overall performance for multiple memcgs running mixed types of workloads. 10. mm: multi-gen LRU: kill switch Adds a kill switch to enable or disable MGLRU at runtime. 11. mm: multi-gen LRU: thrashing prevention 12. mm: multi-gen LRU: debugfs interface Provide userspace with features like thrashing prevention, working set estimation and proactive reclaim. 13. mm: multi-gen LRU: admin guide 14. mm: multi-gen LRU: design doc Add an admin guide and a design doc. Benchmark results ================= Independent lab results ----------------------- Based on the popularity of searches [01] and the memory usage in Google's public cloud, the most popular open-source memory-hungry applications, in alphabetical order, are: Apache Cassandra Memcached Apache Hadoop MongoDB Apache Spark PostgreSQL MariaDB (MySQL) Redis An independent lab evaluated MGLRU with the most widely used benchmark suites for the above applications. They posted 960 data points along with kernel metrics and perf profiles collected over more than 500 hours of total benchmark time. Their final reports show that, with 95% confidence intervals (CIs), the above applications all performed significantly better for at least part of their benchmark matrices. On 5.14: 1. Apache Spark [02] took 95% CIs [9.28, 11.19]% and [12.20, 14.93]% less wall time to sort three billion random integers, respectively, under the medium- and the high-concurrency conditions, when overcommitting memory. There were no statistically significant changes in wall time for the rest of the benchmark matrix. 2. MariaDB [03] achieved 95% CIs [5.24, 10.71]% and [20.22, 25.97]% more transactions per minute (TPM), respectively, under the medium- and the high-concurrency conditions, when overcommitting memory. There were no statistically significant changes in TPM for the rest of the benchmark matrix. 3. Memcached [04] achieved 95% CIs [23.54, 32.25]%, [20.76, 41.61]% and [21.59, 30.02]% more operations per second (OPS), respectively, for sequential access, random access and Gaussian (distribution) access, when THP=always; 95% CIs [13.85, 15.97]% and [23.94, 29.92]% more OPS, respectively, for random access and Gaussian access, when THP=never. There were no statistically significant changes in OPS for the rest of the benchmark matrix. 4. MongoDB [05] achieved 95% CIs [2.23, 3.44]%, [6.97, 9.73]% and [2.16, 3.55]% more operations per second (OPS), respectively, for exponential (distribution) access, random access and Zipfian (distribution) access, when underutilizing memory; 95% CIs [8.83, 10.03]%, [21.12, 23.14]% and [5.53, 6.46]% more OPS, respectively, for exponential access, random access and Zipfian access, when overcommitting memory. On 5.15: 5. Apache Cassandra [06] achieved 95% CIs [1.06, 4.10]%, [1.94, 5.43]% and [4.11, 7.50]% more operations per second (OPS), respectively, for exponential (distribution) access, random access and Zipfian (distribution) access, when swap was off; 95% CIs [0.50, 2.60]%, [6.51, 8.77]% and [3.29, 6.75]% more OPS, respectively, for exponential access, random access and Zipfian access, when swap was on. 6. Apache Hadoop [07] took 95% CIs [5.31, 9.69]% and [2.02, 7.86]% less average wall time to finish twelve parallel TeraSort jobs, respectively, under the medium- and the high-concurrency conditions, when swap was on. There were no statistically significant changes in average wall time for the rest of the benchmark matrix. 7. PostgreSQL [08] achieved 95% CI [1.75, 6.42]% more transactions per minute (TPM) under the high-concurrency condition, when swap was off; 95% CIs [12.82, 18.69]% and [22.70, 46.86]% more TPM, respectively, under the medium- and the high-concurrency conditions, when swap was on. There were no statistically significant changes in TPM for the rest of the benchmark matrix. 8. Redis [09] achieved 95% CIs [0.58, 5.94]%, [6.55, 14.58]% and [11.47, 19.36]% more total operations per second (OPS), respectively, for sequential access, random access and Gaussian (distribution) access, when THP=always; 95% CIs [1.27, 3.54]%, [10.11, 14.81]% and [8.75, 13.64]% more total OPS, respectively, for sequential access, random access and Gaussian access, when THP=never. Our lab results --------------- To supplement the above results, we ran the following benchmark suites on 5.16-rc7 and found no regressions [10]. fs_fio_bench_hdd_mq pft fs_lmbench pgsql-hammerdb fs_parallelio redis fs_postmark stream hackbench sysbenchthread kernbench tpcc_spark memcached unixbench multichase vm-scalability mutilate will-it-scale nginx [01] https://trends.google.com [02] https://lore.kernel.org/r/20211102002002.92051-1-bot@edi.works/ [03] https://lore.kernel.org/r/20211009054315.47073-1-bot@edi.works/ [04] https://lore.kernel.org/r/20211021194103.65648-1-bot@edi.works/ [05] https://lore.kernel.org/r/20211109021346.50266-1-bot@edi.works/ [06] https://lore.kernel.org/r/20211202062806.80365-1-bot@edi.works/ [07] https://lore.kernel.org/r/20211209072416.33606-1-bot@edi.works/ [08] https://lore.kernel.org/r/20211218071041.24077-1-bot@edi.works/ [09] https://lore.kernel.org/r/20211122053248.57311-1-bot@edi.works/ [10] https://lore.kernel.org/r/20220104202247.2903702-1-yuzhao@google.com/ Read-world applications ======================= Third-party testimonials ------------------------ Konstantin reported [11]: I have Archlinux with 8G RAM + zswap + swap. While developing, I have lots of apps opened such as multiple LSP-servers for different langs, chats, two browsers, etc... Usually, my system gets quickly to a point of SWAP-storms, where I have to kill LSP-servers, restart browsers to free memory, etc, otherwise the system lags heavily and is barely usable. 1.5 day ago I migrated from 5.11.15 kernel to 5.12 + the LRU patchset, and I started up by opening lots of apps to create memory pressure, and worked for a day like this. Till now I had not a single SWAP-storm, and mind you I got 3.4G in SWAP. I was never getting to the point of 3G in SWAP before without a single SWAP-storm. Vaibhav from IBM reported [12]: In a synthetic MongoDB Benchmark, seeing an average of ~19% throughput improvement on POWER10(Radix MMU + 64K Page Size) with MGLRU patches on top of 5.16 kernel for MongoDB + YCSB across three different request distributions, namely, Exponential, Uniform and Zipfan. Shuang from U of Rochester reported [13]: With the MGLRU, fio achieved 95% CIs [38.95, 40.26]%, [4.12, 6.64]% and [9.26, 10.36]% higher throughput, respectively, for random access, Zipfian (distribution) access and Gaussian (distribution) access, when the average number of jobs per CPU is 1; 95% CIs [42.32, 49.15]%, [9.44, 9.89]% and [20.99, 22.86]% higher throughput, respectively, for random access, Zipfian access and Gaussian access, when the average number of jobs per CPU is 2. Daniel from Michigan Tech reported [14]: With Memcached allocating ~100GB of byte-addressable Optante, performance improvement in terms of throughput (measured as queries per second) was about 10% for a series of workloads. Large-scale deployments ----------------------- We've rolled out MGLRU to tens of millions of ChromeOS users and about a million Android users. Google's fleetwide profiling [15] shows an overall 40% decrease in kswapd CPU usage, in addition to improvements in other UX metrics, e.g., an 85% decrease in the number of low-memory kills at the 75th percentile and an 18% decrease in app launch time at the 50th percentile. The downstream kernels that have been using MGLRU include: 1. Android [16] 2. Arch Linux Zen [17] 3. Armbian [18] 4. ChromeOS [19] 5. Liquorix [20] 6. OpenWrt [21] 7. post-factum [22] 8. XanMod [23] [11] https://lore.kernel.org/r/140226722f2032c86301fbd326d91baefe3d7d23.camel@yandex.ru/ [12] https://lore.kernel.org/r/87czj3mux0.fsf@vajain21.in.ibm.com/ [13] https://lore.kernel.org/r/20220105024423.26409-1-szhai2@cs.rochester.edu/ [14] https://lore.kernel.org/r/CA+4-3vksGvKd18FgRinxhqHetBS1hQekJE2gwco8Ja-bJWKtFw@mail.gmail.com/ [15] https://dl.acm.org/doi/10.1145/2749469.2750392 [16] https://android.com [17] https://archlinux.org [18] https://armbian.com [19] https://chromium.org [20] https://liquorix.net [21] https://openwrt.org [22] https://codeberg.org/pf-kernel [23] https://xanmod.org Summary ======= The facts are: 1. The independent lab results and the real-world applications indicate substantial improvements; there are no known regressions. 2. Thrashing prevention, working set estimation and proactive reclaim work out of the box; there are no equivalent solutions. 3. There is a lot of new code; no smaller changes have been demonstrated similar effects. Our options, accordingly, are: 1. Given the amount of evidence, the reported improvements will likely materialize for a wide range of workloads. 2. Gauging the interest from the past discussions, the new features will likely be put to use for both personal computers and data centers. 3. Based on Google's track record, the new code will likely be well maintained in the long term. It'd be more difficult if not impossible to achieve similar effects with other approaches. This patch (of 14): Some architectures automatically set the accessed bit in PTEs, e.g., x86 and arm64 v8.2. On architectures that do not have this capability, clearing the accessed bit in a PTE usually triggers a page fault following the TLB miss of this PTE (to emulate the accessed bit). Being aware of this capability can help make better decisions, e.g., whether to spread the work out over a period of time to reduce bursty page faults when trying to clear the accessed bit in many PTEs. Note that theoretically this capability can be unreliable, e.g., hotplugged CPUs might be different from builtin ones. Therefore it should not be used in architecture-independent code that involves correctness, e.g., to determine whether TLB flushes are required (in combination with the accessed bit). Link: https://lkml.kernel.org/r/20220918080010.2920238-1-yuzhao@google.com Link: https://lkml.kernel.org/r/20220918080010.2920238-2-yuzhao@google.com Signed-off-by: Yu Zhao Reviewed-by: Barry Song Acked-by: Brian Geffon Acked-by: Jan Alexander Steffens (heftig) Acked-by: Oleksandr Natalenko Acked-by: Steven Barrett Acked-by: Suleiman Souhlal Acked-by: Will Deacon Tested-by: Daniel Byrne Tested-by: Donald Carr Tested-by: Holger Hoffstätte Tested-by: Konstantin Kharlamov Tested-by: Shuang Zhai Tested-by: Sofia Trinh Tested-by: Vaibhav Jain Cc: Andi Kleen Cc: Aneesh Kumar K.V Cc: Catalin Marinas Cc: Dave Hansen Cc: Hillf Danton Cc: Jens Axboe Cc: Johannes Weiner Cc: Jonathan Corbet Cc: Linus Torvalds Cc: linux-arm-kernel@lists.infradead.org Cc: Matthew Wilcox Cc: Mel Gorman Cc: Michael Larabel Cc: Michal Hocko Cc: Mike Rapoport Cc: Peter Zijlstra Cc: Tejun Heo Cc: Vlastimil Babka Cc: Miaohe Lin Cc: Mike Rapoport Cc: Qi Zheng Signed-off-by: Andrew Morton --- arch/arm64/include/asm/pgtable.h | 14 ++------------ arch/x86/include/asm/pgtable.h | 6 +++--- include/linux/pgtable.h | 13 +++++++++++++ mm/memory.c | 14 +------------- 4 files changed, 19 insertions(+), 28 deletions(-) --- a/arch/arm64/include/asm/pgtable.h +++ b/arch/arm64/include/asm/pgtable.h @@ -1005,23 +1005,13 @@ static inline void update_mmu_cache(stru * page after fork() + CoW for pfn mappings. We don't always have a * hardware-managed access flag on arm64. */ -static inline bool arch_faults_on_old_pte(void) -{ - WARN_ON(preemptible()); - - return !cpu_has_hw_af(); -} -#define arch_faults_on_old_pte arch_faults_on_old_pte +#define arch_has_hw_pte_young cpu_has_hw_af /* * Experimentally, it's cheap to set the access flag in hardware and we * benefit from prefaulting mappings as 'old' to start with. */ -static inline bool arch_wants_old_prefaulted_pte(void) -{ - return !arch_faults_on_old_pte(); -} -#define arch_wants_old_prefaulted_pte arch_wants_old_prefaulted_pte +#define arch_wants_old_prefaulted_pte cpu_has_hw_af #endif /* !__ASSEMBLY__ */ --- a/arch/x86/include/asm/pgtable.h +++ b/arch/x86/include/asm/pgtable.h @@ -1397,10 +1397,10 @@ static inline bool arch_has_pfn_modify_c return boot_cpu_has_bug(X86_BUG_L1TF); } -#define arch_faults_on_old_pte arch_faults_on_old_pte -static inline bool arch_faults_on_old_pte(void) +#define arch_has_hw_pte_young arch_has_hw_pte_young +static inline bool arch_has_hw_pte_young(void) { - return false; + return true; } #endif /* __ASSEMBLY__ */ --- a/include/linux/pgtable.h +++ b/include/linux/pgtable.h @@ -259,6 +259,19 @@ static inline int pmdp_clear_flush_young #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ #endif +#ifndef arch_has_hw_pte_young +/* + * Return whether the accessed bit is supported on the local CPU. + * + * This stub assumes accessing through an old PTE triggers a page fault. + * Architectures that automatically set the access bit should overwrite it. + */ +static inline bool arch_has_hw_pte_young(void) +{ + return false; +} +#endif + #ifndef __HAVE_ARCH_PTEP_GET_AND_CLEAR static inline pte_t ptep_get_and_clear(struct mm_struct *mm, unsigned long address, --- a/mm/memory.c +++ b/mm/memory.c @@ -121,18 +121,6 @@ int randomize_va_space __read_mostly = 2; #endif -#ifndef arch_faults_on_old_pte -static inline bool arch_faults_on_old_pte(void) -{ - /* - * Those arches which don't have hw access flag feature need to - * implement their own helper. By default, "true" means pagefault - * will be hit on old pte. - */ - return true; -} -#endif - #ifndef arch_wants_old_prefaulted_pte static inline bool arch_wants_old_prefaulted_pte(void) { @@ -2791,7 +2779,7 @@ static inline int cow_user_page(struct p * On architectures with software "accessed" bits, we would * take a double page fault, so mark it accessed here. */ - if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) { + if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) { pte_t entry; vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);