From: SeongJae Park <sjpark@xxxxxxxxx> Introduction ============ Memory management decisions can be improved if finer data access information is available. However, because such finer information usually comes with higher overhead, most systems including Linux forgives the potential improvement and rely on only coarse information or some light-weight heuristics. The pseudo-LRU and the aggressive THP promotions are such examples. A number of experimental data access pattern awared memory management optimizations (refer to 'Appendix A' for more details) say the sacrifices are huge. However, none of those has successfully adopted to Linux kernel mainly due to the absence of a scalable and efficient data access monitoring mechanism. Refer to 'Appendix B' to see the limitations of existing memory monitoring mechanisms. DAMON is a data access monitoring subsystem for the problem. It is 1) accurate enough to be used for the DRAM level memory management (a straightforward DAMON-based optimization achieved up to 2.55x speedup), 2) light-weight enough to be applied online (compared to a straightforward access monitoring scheme, DAMON is up to 94.242.42x lighter) and 3) keeps predefined upper-bound overhead regardless of the size of target workloads (thus scalable). Refer to 'Appendix C' if you interested in how it is possible. DAMON has mainly designed for the kernel's memory management mechanisms. However, because it is implemented as a standalone kernel module and provides several interfaces, it can be used by a wide range of users including kernel space programs, user space programs, programmers, and administrators. DAMON is now supporting the monitoring only, but it will also provide simple and convenient data access pattern awared memory managements by itself. Refer to 'Appendix D' for more detailed expected usages of DAMON. Visualized Outputs of DAMON =========================== For intuitively understanding of DAMON, I made web pages[1-8] showing the visualized dynamic data access pattern of various realistic workloads, which I picked up from PARSEC3 and SPLASH-2X bechmark suites. The figures are generated using the user space tool in 10th patch of this patchset. There are pages showing the heatmap format dynamic access pattern of each workload for heap area[1], mmap()-ed area[2], and stack[3] area. I splitted the entire address space to the three area because there are huge unmapped regions between the areas. You can also show how the dynamic working set size of each workload is distributed[4], and how it is chronologically changing[5]. The most important characteristic of DAMON is its promise of the upperbound of the monitoring overhead. To show whether DAMON keeps the promise well, I visualized the number of monitoring operations required for each 5 milliseconds, which is configured to not exceed 1000. You can show the distribution of the numbers[6] and how it changes chronologically[7]. [1] https://damonitor.github.io/reports/latest/by_image/heatmap.0.png.html [2] https://damonitor.github.io/reports/latest/by_image/heatmap.1.png.html [3] https://damonitor.github.io/reports/latest/by_image/heatmap.2.png.html [4] https://damonitor.github.io/reports/latest/by_image/wss_sz.png.html [5] https://damonitor.github.io/reports/latest/by_image/wss_time.png.html [6] https://damonitor.github.io/reports/latest/by_image/nr_regions_sz.png.html [7] https://damonitor.github.io/reports/latest/by_image/nr_regions_time.png.html Data Access Monitoring-based Operation Schemes ============================================== As 'Appendix D' describes, DAMON can be used for data access monitoring-based operation schemes (DAMOS). RFC patchsets for DAMOS are already available (https://lore.kernel.org/linux-mm/20200218085309.18346-1-sjpark@xxxxxxxxxx/). By applying a very simple scheme for THP promotion/demotion with a latest version of the patchset (not posted yet), DAMON achieved 18x lower memory space overhead compared to THP while preserving about 50% of the THP performance benefit with SPLASH-2X benchmark suite. The detailed setup and number will be posted soon with the next RFC patchset for DAMOS. The posting is currently scheduled for tomorrow. Frequently Asked Questions ========================== Q: Why DAMON is not integrated with perf? A: From the perspective of perf like profilers, DAMON can be thought of as a data source in kernel, like the tracepoints, the pressure stall information (psi), or the idle page tracking. Thus, it is easy to integrate DAMON with the profilers. However, this patchset doesn't provide a fancy perf integration because current step of DAMON development is focused on its core logic only. That said, DAMON already provides two interfaces for user space programs, which based on debugfs and tracepoint, respectively. Using the tracepoint interface, you can use DAMON with perf. This patchset also provides a debugfs interface based user space tool for DAMON. It can be used to record, visualize, and analyze data access patterns of target processes in a convenient way. Q: Why a new module, instead of extending perf or other tools? A: First, DAMON aims to be used by other programs including the kernel. Therefore, having dependency to specific tools like perf is not desirable. Second, because it need to be lightweight as much as possible so that it can be used online, any unnecessary overhead such as kernel - user space context switching cost should be avoided. These are the two most biggest reasons why DAMON is implemented in the kernel space. The idle page tracking subsystem would be the kernel module that most seems similar to DAMON. However, its own interface is not compatible with DAMON. Also, the internal implementation of it has no common part to be reused by DAMON. Q: Can 'perf mem' provide the data required for DAMON? A: On the systems supporting 'perf mem', yes. DAMON is using the PTE Accessed bits in low level. Other H/W or S/W features that can be used for the purpose could be used. However, as explained with above question, DAMON need to be implemented in the kernel space. Evaluations =========== A prototype of DAMON has evaluated on an Intel Xeon E7-8837 machine using 20 benchmarks that picked from SPEC CPU 2006, NAS, Tensorflow Benchmark, SPLASH-2X, and PARSEC 3 benchmark suite. Nonethless, this section provides only summary of the results. For more detail, please refer to the slides used for the introduction of DAMON at the Linux Plumbers Conference 2019[1] or the MIDDLEWARE'19 industrial track paper[2]. Quality ------- We first traced and visualized the data access pattern of each workload. We were able to confirm that the visualized results are reasonably accurate by manually comparing those with the source code of the workloads. To see the usefulness of the monitoring, we optimized 9 memory intensive workloads among them for memory pressure situations using the DAMON outputs. In detail, we identified frequently accessed memory regions in each workload based on the DAMON results and protected them with ``mlock()`` system calls. The optimized versions consistently show speedup (2.55x in best case, 1.65x in average) under memory pressure. Overhead -------- We also measured the overhead of DAMON. It was not only under the upperbound we set, but was much lower (0.6 percent of the bound in best case, 13.288 percent of the bound in average). This reduction of the overhead is mainly resulted from its core mechanism called adaptive regions adjustment. Refer to 'Appendix D' for more detail about the mechanism. We also compared the overhead of DAMON with that of a straightforward periodic access check-based monitoring. DAMON's overhead was smaller than it by 94,242.42x in best case, 3,159.61x in average. The latest version of DAMON running with its default configuration consumes only up to 1% of CPU time when applied to realistic workloads in PARSEC3 and SPLASH-2X and makes no visible slowdown to the target processes. References ========== Prototypes of DAMON have introduced by an LPC kernel summit track talk[1] and two academic papers[2,3]. Please refer to those for more detailed information, especially the evaluations. The latest version of the patchsets has also introduced by an LWN artice[4]. [1] SeongJae Park, Tracing Data Access Pattern with Bounded Overhead and Best-effort Accuracy. In The Linux Kernel Summit, September 2019. https://linuxplumbersconf.org/event/4/contributions/548/ [2] SeongJae Park, Yunjae Lee, Heon Y. Yeom, Profiling Dynamic Data Access Patterns with Controlled Overhead and Quality. In 20th ACM/IFIP International Middleware Conference Industry, December 2019. https://dl.acm.org/doi/10.1145/3366626.3368125 [3] SeongJae Park, Yunjae Lee, Yunhee Kim, Heon Y. Yeom, Profiling Dynamic Data Access Patterns with Bounded Overhead and Accuracy. In IEEE International Workshop on Foundations and Applications of Self- Systems (FAS 2019), June 2019. [4] Jonathan Corbet, Memory-management optimization with DAMON. In Linux Weekly News (LWN), Feb 2020. https://lwn.net/Articles/812707/ Sequence Of Patches =================== The patches are organized in the following sequence. The first patch introduces DAMON module, it's data structures, and data structure related common functions. Following three patches (2nd to 4th) implement the core logics of DAMON, namely regions based sampling, adaptive regions adjustment, and dynamic memory mapping chage adoption, one by one. Following five patches are for low level users of DAMON. The 5th patch implements callbacks for each of monitoring steps so that users can do whatever they want with the access patterns. The 6th one implements recording of access patterns in DAMON for better convenience and efficiency. Each of next three patches (7th to 9th) respectively adds a programmable interface for other kernel code, a debugfs interface for privileged people and/or programs in user space, and a tracepoint for other tracepoints supporting tracers such as perf. Two patches for high level users of DAMON follows. To provide a minimal reference to the debugfs interface and for high level use/tests of the DAMON, the next patch (10th) implements an user space tool. The 11th patch adds a document for administrators of DAMON. Next two patches are for tests. The 12th and 13th patches provide unit tests (based on kunit) and user space tests (based on kselftest) respectively. Finally, the last patch (14th) updates the MAINTAINERS file. The patches are based on the v5.5. You can also clone the complete git tree: $ git clone git://github.com/sjp38/linux -b damon/patches/v6 The web is also available: https://github.com/sjp38/linux/releases/tag/damon/patches/v6 Patch History ============= Changes from v5 (https://lore.kernel.org/linux-mm/20200217103110.30817-1-sjpark@xxxxxxxxxx/) - Fix minor bugs (sampling, record attributes, debugfs and user space tool) - selftests: Add debugfs interface tests for the bugs - Modify the user space tool to use its self default values for parameters - Fix pmg huge page access check Changes from v4 (https://lore.kernel.org/linux-mm/20200210144812.26845-1-sjpark@xxxxxxxxxx/) - Add 'Reviewed-by' for the kunit tests patch (Brendan Higgins) - Make the unit test to depedns on 'DAMON=y' (Randy Dunlap and kbuild bot) Reported-by: kbuild test robot <lkp@xxxxxxxxx> - Fix m68k module build issue Reported-by: kbuild test robot <lkp@xxxxxxxxx> - Add selftests - Seperate patches for low level users from core logics for better reading - Clean up debugfs interface - Trivial nitpicks Changes from v3 (https://lore.kernel.org/linux-mm/20200204062312.19913-1-sj38.park@xxxxxxxxx/) - Fix i386 build issue Reported-by: kbuild test robot <lkp@xxxxxxxxx> - Increase the default size of the monitoring result buffer to 1 MiB - Fix misc bugs in debugfs interface Changes from v2 (https://lore.kernel.org/linux-mm/20200128085742.14566-1-sjpark@xxxxxxxxxx/) - Move MAINTAINERS changes to last commit (Brendan Higgins) - Add descriptions for kunittest: why not only entire mappings and what the 4 input sets are trying to test (Brendan Higgins) - Remove 'kdamond_need_stop()' test (Brendan Higgins) - Discuss about the 'perf mem' and DAMON (Peter Zijlstra) - Make CV clearly say what it actually does (Peter Zijlstra) - Answer why new module (Qian Cai) - Diable DAMON by default (Randy Dunlap) - Change the interface: Seperate recording attributes (attrs, record, rules) and allow multiple kdamond instances - Implement kernel API interface Changes from v1 (https://lore.kernel.org/linux-mm/20200120162757.32375-1-sjpark@xxxxxxxxxx/) - Rebase on v5.5 - Add a tracepoint for integration with other tracers (Kirill A. Shutemov) - document: Add more description for the user space tool (Brendan Higgins) - unittest: Improve readability (Brendan Higgins) - unittest: Use consistent name and helpers function (Brendan Higgins) - Update PG_Young to avoid reclaim logic interference (Yunjae Lee) Changes from RFC (https://lore.kernel.org/linux-mm/20200110131522.29964-1-sjpark@xxxxxxxxxx/) - Specify an ambiguous plan of access pattern based mm optimizations - Support loadable module build - Cleanup code SeongJae Park (14): mm: Introduce Data Access MONitor (DAMON) mm/damon: Implement region based sampling mm/damon: Adaptively adjust regions mm/damon: Apply dynamic memory mapping changes mm/damon: Implement callbacks mm/damon: Implement access pattern recording mm/damon: Implement kernel space API mm/damon: Add debugfs interface mm/damon: Add a tracepoint for result writing tools: Add a minimal user-space tool for DAMON Documentation/admin-guide/mm: Add a document for DAMON mm/damon: Add kunit tests mm/damon: Add user selftests MAINTAINERS: Update for DAMON .../admin-guide/mm/data_access_monitor.rst | 414 +++++ Documentation/admin-guide/mm/index.rst | 1 + MAINTAINERS | 12 + include/linux/damon.h | 71 + include/trace/events/damon.h | 32 + mm/Kconfig | 23 + mm/Makefile | 1 + mm/damon-test.h | 604 +++++++ mm/damon.c | 1427 +++++++++++++++++ mm/page_ext.c | 1 + tools/damon/.gitignore | 1 + tools/damon/_dist.py | 36 + tools/damon/bin2txt.py | 64 + tools/damon/damo | 37 + tools/damon/heats.py | 358 +++++ tools/damon/nr_regions.py | 89 + tools/damon/record.py | 212 +++ tools/damon/report.py | 45 + tools/damon/wss.py | 95 ++ tools/testing/selftests/damon/Makefile | 7 + .../selftests/damon/_chk_dependency.sh | 28 + tools/testing/selftests/damon/_chk_record.py | 89 + .../testing/selftests/damon/debugfs_attrs.sh | 139 ++ .../testing/selftests/damon/debugfs_record.sh | 50 + 24 files changed, 3836 insertions(+) create mode 100644 Documentation/admin-guide/mm/data_access_monitor.rst create mode 100644 include/linux/damon.h create mode 100644 include/trace/events/damon.h create mode 100644 mm/damon-test.h create mode 100644 mm/damon.c create mode 100644 tools/damon/.gitignore create mode 100644 tools/damon/_dist.py create mode 100644 tools/damon/bin2txt.py create mode 100755 tools/damon/damo create mode 100644 tools/damon/heats.py create mode 100644 tools/damon/nr_regions.py create mode 100644 tools/damon/record.py create mode 100644 tools/damon/report.py create mode 100644 tools/damon/wss.py create mode 100644 tools/testing/selftests/damon/Makefile create mode 100644 tools/testing/selftests/damon/_chk_dependency.sh create mode 100644 tools/testing/selftests/damon/_chk_record.py create mode 100755 tools/testing/selftests/damon/debugfs_attrs.sh create mode 100755 tools/testing/selftests/damon/debugfs_record.sh -- 2.17.1 ============================= 8< ====================================== Appendix A: Related Works ========================= There are a number of researches[1,2,3,4,5,6] optimizing memory management mechanisms based on the actual memory access patterns that shows impressive results. However, most of those has no deep consideration about the monitoring of the accesses itself. Some of those focused on the overhead of the monitoring, but does not consider the accuracy scalability[6] or has additional dependencies[7]. Indeed, one recent research[5] about the proactive reclamation has also proposed[8] to the kernel community but the monitoring overhead was considered a main problem. [1] Subramanya R Dulloor, Amitabha Roy, Zheguang Zhao, Narayanan Sundaram, Nadathur Satish, Rajesh Sankaran, Jeff Jackson, and Karsten Schwan. 2016. Data tiering in heterogeneous memory systems. In Proceedings of the 11th European Conference on Computer Systems (EuroSys). ACM, 15. [2] Youngjin Kwon, Hangchen Yu, Simon Peter, Christopher J Rossbach, and Emmett Witchel. 2016. Coordinated and efficient huge page management with ingens. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI). 705–721. [3] Harald Servat, Antonio J Peña, Germán Llort, Estanislao Mercadal, HansChristian Hoppe, and Jesús Labarta. 2017. Automating the application data placement in hybrid memory systems. In 2017 IEEE International Conference on Cluster Computing (CLUSTER). IEEE, 126–136. [4] Vlad Nitu, Boris Teabe, Alain Tchana, Canturk Isci, and Daniel Hagimont. 2018. Welcome to zombieland: practical and energy-efficient memory disaggregation in a datacenter. In Proceedings of the 13th European Conference on Computer Systems (EuroSys). ACM, 16. [5] Andres Lagar-Cavilla, Junwhan Ahn, Suleiman Souhlal, Neha Agarwal, Radoslaw Burny, Shakeel Butt, Jichuan Chang, Ashwin Chaugule, Nan Deng, Junaid Shahid, Greg Thelen, Kamil Adam Yurtsever, Yu Zhao, and Parthasarathy Ranganathan. 2019. Software-Defined Far Memory in Warehouse-Scale Computers. In Proceedings of the 24th International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS). ACM, New York, NY, USA, 317–330. DOI:https://doi.org/10.1145/3297858.3304053 [6] Carl Waldspurger, Trausti Saemundsson, Irfan Ahmad, and Nohhyun Park. 2017. Cache Modeling and Optimization using Miniature Simulations. In 2017 USENIX Annual Technical Conference (ATC). USENIX Association, Santa Clara, CA, 487–498. https://www.usenix.org/conference/atc17/technical-sessions/ [7] Haojie Wang, Jidong Zhai, Xiongchao Tang, Bowen Yu, Xiaosong Ma, and Wenguang Chen. 2018. Spindle: Informed Memory Access Monitoring. In 2018 USENIX Annual Technical Conference (ATC). USENIX Association, Boston, MA, 561–574. https://www.usenix.org/conference/atc18/presentation/wang-haojie [8] Jonathan Corbet. 2019. Proactively reclaiming idle memory. (2019). https://lwn.net/Articles/787611/. Appendix B: Limitations of Other Access Monitoring Techniques ============================================================= The memory access instrumentation techniques which are applied to many tools such as Intel PIN is essential for correctness required cases such as memory access bug detections or cache level optimizations. However, those usually incur exceptionally high overhead which is unacceptable. Periodic access checks based on access counting features (e.g., PTE Accessed bits or PG_Idle flags) can reduce the overhead. It sacrifies some of the quality but it's still ok to many of this domain. However, the overhead arbitrarily increase as the size of the target workload grows. Miniature-like static region based sampling can set the upperbound of the overhead, but it will now decrease the quality of the output as the size of the workload grows. DAMON is another solution that overcomes the limitations. It is 1) accurate enough for this domain, 2) light-weight so that it can be applied online, and 3) allow users to set the upper-bound of the overhead, regardless of the size of target workloads. It is implemented as a simple and small kernel module to support various users in both of the user space and the kernel space. Refer to 'Evaluations' section below for detailed performance of DAMON. For the goals, DAMON utilizes its two core mechanisms, which allows lightweight overhead and high quality of output, repectively. To show how DAMON promises those, refer to 'Mechanisms of DAMON' section below. Appendix C: Mechanisms of DAMON =============================== Basic Access Check ------------------ DAMON basically reports what pages are how frequently accessed. The report is passed to users in binary format via a ``result file`` which users can set it's path. Note that the frequency is not an absolute number of accesses, but a relative frequency among the pages of the target workloads. Users can also control the resolution of the reports by setting two time intervals, ``sampling interval`` and ``aggregation interval``. In detail, DAMON checks access to each page per ``sampling interval``, aggregates the results (counts the number of the accesses to each page), and reports the aggregated results per ``aggregation interval``. For the access check of each page, DAMON uses the Accessed bits of PTEs. This is thus similar to the previously mentioned periodic access checks based mechanisms, which overhead is increasing as the size of the target process grows. Region Based Sampling --------------------- To avoid the unbounded increase of the overhead, DAMON groups a number of adjacent pages that assumed to have same access frequencies into a region. As long as the assumption (pages in a region have same access frequencies) is kept, only one page in the region is required to be checked. Thus, for each ``sampling interval``, DAMON randomly picks one page in each region and clears its Accessed bit. After one more ``sampling interval``, DAMON reads the Accessed bit of the page and increases the access frequency of the region if the bit has set meanwhile. Therefore, the monitoring overhead is controllable by setting the number of regions. DAMON allows users to set the minimal and maximum number of regions for the trade-off. Except the assumption, this is almost same with the above-mentioned miniature-like static region based sampling. In other words, this scheme cannot preserve the quality of the output if the assumption is not guaranteed. Adaptive Regions Adjustment --------------------------- At the beginning of the monitoring, DAMON constructs the initial regions by evenly splitting the memory mapped address space of the process into the user-specified minimal number of regions. In this initial state, the assumption is normally not kept and thus the quality could be low. To keep the assumption as much as possible, DAMON adaptively merges and splits each region. For each ``aggregation interval``, it compares the access frequencies of adjacent regions and merges those if the frequency difference is small. Then, after it reports and clears the aggregated access frequency of each region, it splits each region into two regions if the total number of regions is smaller than the half of the user-specified maximum number of regions. In this way, DAMON provides its best-effort quality and minimal overhead while keeping the bounds users set for their trade-off. Applying Dynamic Memory Mappings -------------------------------- Only a number of small parts in the super-huge virtual address space of the processes is mapped to physical memory and accessed. Thus, tracking the unmapped address regions is just wasteful. However, tracking every memory mapping change might incur an overhead. For the reason, DAMON applies the dynamic memory mapping changes to the tracking regions only for each of an user-specified time interval (``regions update interval``). Appendix D: Expected Use-cases ============================== A straightforward usecase of DAMON would be the program behavior analysis. With the DAMON output, users can confirm whether the program is running as intended or not. This will be useful for debuggings and tests of design points. The monitored results can also be useful for counting the dynamic working set size of workloads. For the administration of memory overcommitted systems or selection of the environments (e.g., containers providing different amount of memory) for your workloads, this will be useful. If you are a programmer, you can optimize your program by managing the memory based on the actual data access pattern. For example, you can identify the dynamic hotness of your data using DAMON and call ``mlock()`` to keep your hot data in DRAM, or call ``madvise()`` with ``MADV_PAGEOUT`` to proactively reclaim cold data. Even though your program is guaranteed to not encounter memory pressure, you can still improve the performance by applying the DAMON outputs for call of ``MADV_HUGEPAGE`` and ``MADV_NOHUGEPAGE``. More creative optimizations would be possible. Our evaluations of DAMON includes a straightforward optimization using the ``mlock()``. Please refer to the below Evaluation section for more detail. As DAMON incurs very low overhead, such optimizations can be applied not only offline, but also online. Also, there is no reason to limit such optimizations to the user space. Several parts of the kernel's memory management mechanisms could be also optimized using DAMON. The reclamation, the THP (de)promotion decisions, and the compaction would be such a candidates. DAMON will continue its development to be highly optimized for the online/in-kernel uses. A Future Plan: Data Access Monitoring-based Operation Schemes ------------------------------------------------------------- As described in the above section, DAMON could be helpful for actual access based memory management optimizations. Nevertheless, users who want to do such optimizations should run DAMON, read the traced data (either online or offline), analyze it, plan a new memory management scheme, and apply the new scheme by themselves. It must be easier than the past, but could still require some level of efforts. In its next development stage, DAMON will reduce some of such efforts by allowing users to specify some access based memory management rules for their specific processes. Because this is just a plan, the specific interface is not fixed yet, but for example, users will be allowed to write their desired memory management rules to a special file in a DAMON specific format. The rules will be something like 'if a memory region of size in a range is keeping a range of hotness for more than a duration, apply specific memory management rule using madvise() or mlock() to the region'. For example, we can imagine rules like below: # format is: <min/max size> <min/max frequency (0-99)> <duration> <action> # if a region of a size keeps a very high access frequency for more than # 100ms, lock the region in the main memory (call mlock()). But, if the # region is larger than 500 MiB, skip it. The exception might be helpful # if the system has only, say, 600 MiB of DRAM, a region of size larger # than 600 MiB cannot be locked in the DRAM at all. na 500M 90 99 100ms mlock # if a region keeps a high access frequency for more than 100ms, put the # region on the head of the LRU list (call madvise() with MADV_WILLNEED). na na 80 90 100ms madv_willneed # if a region keeps a low access frequency for more than 100ms, put the # region on the tail of the LRU list (call madvise() with MADV_COLD). na na 10 20 100ms madv_cold # if a region keeps a very low access frequency for more than 100ms, swap # out the region immediately (call madvise() with MADV_PAGEOUT). na na 0 10 100ms madv_pageout # if a region of a size bigger than 2MB keeps a very high access frequency # for more than 100ms, let the region to use huge pages (call madvise() # with MADV_HUGEPAGE). 2M na 90 99 100ms madv_hugepage # If a regions of a size bigger than > 2MB keeps no high access frequency # for more than 100ms, avoid the region from using huge pages (call # madvise() with MADV_NOHUGEPAGE). 2M na 0 25 100ms madv_nohugepage An RFC patchset for this is available: https://lore.kernel.org/linux-mm/20200218085309.18346-1-sjpark@xxxxxxxxxx/
