| Message ID | 20190404020046.32741-1-zi.yan@sent.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Accelerate page migration and use memcg for PMEM management | expand |
On Wed 03-04-19 19:00:21, Zi Yan wrote: > From: Zi Yan <ziy@nvidia.com> > > Thanks to Dave Hansen's patches, which make PMEM as part of memory as NUMA nodes. > How to use PMEM along with normal DRAM remains an open problem. There are > several patchsets posted on the mailing list, proposing to use page migration to > move pages between PMEM and DRAM using Linux page replacement policy [1,2,3]. > There are some important problems not addressed in these patches: > 1. The page migration in Linux does not provide high enough throughput for us to > fully exploit PMEM or other use cases. > 2. Linux page replacement is running too infrequent to distinguish hot and cold > pages. [...] > 33 files changed, 4261 insertions(+), 162 deletions(-) For a patch _this_ large you should really start with a real world usecasing hitting bottlenecks with the current implementation. Should microbenchmarks can trigger bottlenecks much easier but do real application do the same? Please give us some numbers.
On 4/3/19 7:00 PM, Zi Yan wrote: > From: Zi Yan <ziy@nvidia.com> > > Thanks to Dave Hansen's patches, which make PMEM as part of memory as NUMA nodes. > How to use PMEM along with normal DRAM remains an open problem. There are > several patchsets posted on the mailing list, proposing to use page migration to > move pages between PMEM and DRAM using Linux page replacement policy [1,2,3]. > There are some important problems not addressed in these patches: > 1. The page migration in Linux does not provide high enough throughput for us to > fully exploit PMEM or other use cases. > 2. Linux page replacement is running too infrequent to distinguish hot and cold > pages. > > I am trying to attack the problems with this patch series. This is not a final > solution, but I would like to gather more feedback and comments from the mailing > list. > > Page migration throughput problem > ==== > > For example, in my recent email [4], I gave the page migration throughput numbers > for different page migrations, none of which can achieve > 2.5GB/s throughput > (the throughput is measured around kernel functions: migrate_pages() and > migrate_page_copy()): > > | migrate_pages() | migrate_page_copy() > migrating single 4KB page: | 0.312GB/s | 1.385GB/s > migrating 512 4KB pages: | 0.854GB/s | 1.983GB/s > migrating single 2MB THP: | 2.387GB/s | 2.481GB/s > > In reality, microbenchmarks show that Intel PMEM can provide ~65GB/s read > throughput and ~16GB/s write throughput [5], which are much higher than > the throughput achieved by Linux page migration. > > In addition, it is also desirable to use page migration to move data > between high-bandwidth memory and DRAM, like IBM Summit, which exposes > high-performance GPU memories as NUMA nodes [6]. This requires even higher page > migration throughput. > > In this patch series, I propose four different ways of improving page migration > throughput (mostly on 2MB THP migration): > 1. multi-threaded page migration: Patch 03 to 06. > 2. DMA-based (using Intel IOAT DMA) page migration: Patch 07 and 08. > 3. concurrent (batched) page migration: Patch 09, 10, and 11. > 4. exchange pages: Patch 12 to 17. (This is a repost of part of [7]) > > Here are some throughput numbers showing clear throughput improvements on > a two-socket NUMA machine with two Xeon E5-2650 v3 @ 2.30GHz and a 19.2GB/s > bandwidth QPI link (the same machine as mentioned in [4]): > > | migrate_pages() | migrate_page_copy() > => migrating single 2MB THP | 2.387GB/s | 2.481GB/s > 2-thread single THP migration | 3.478GB/s | 3.704GB/s > 4-thread single THP migration | 5.474GB/s | 6.054GB/s > 8-thread single THP migration | 7.846GB/s | 9.029GB/s > 16-thread single THP migration | 7.423GB/s | 8.464GB/s > 16-ch. DMA single THP migration | 4.322GB/s | 4.536GB/s > > 2-thread 16-THP migration | 3.610GB/s | 3.838GB/s > 2-thread 16-THP batched migration | 4.138GB/s | 4.344GB/s > 4-thread 16-THP migration | 6.385GB/s | 7.031GB/s > 4-thread 16-THP batched migration | 7.382GB/s | 8.072GB/s > 8-thread 16-THP migration | 8.039GB/s | 9.029GB/s > 8-thread 16-THP batched migration | 9.023GB/s | 10.056GB/s > 16-thread 16-THP migration | 8.137GB/s | 9.137GB/s > 16-thread 16-THP batched migration | 9.907GB/s | 11.175GB/s > > 1-thread 16-THP exchange | 4.135GB/s | 4.225GB/s > 2-thread 16-THP batched exchange | 7.061GB/s | 7.325GB/s > 4-thread 16-THP batched exchange | 9.729GB/s | 10.237GB/s > 8-thread 16-THP batched exchange | 9.992GB/s | 10.533GB/s > 16-thread 16-THP batched exchange | 9.520GB/s | 10.056GB/s > > => migrating 512 4KB pages | 0.854GB/s | 1.983GB/s > 1-thread 512-4KB batched exchange | 1.271GB/s | 3.433GB/s > 2-thread 512-4KB batched exchange | 1.240GB/s | 3.190GB/s > 4-thread 512-4KB batched exchange | 1.255GB/s | 3.823GB/s > 8-thread 512-4KB batched exchange | 1.336GB/s | 3.921GB/s > 16-thread 512-4KB batched exchange | 1.334GB/s | 3.897GB/s > > Concerns were raised on how to avoid CPU resource competition between > page migration and user applications and have power awareness. > Daniel Jordan recently posted a multi-threaded ktask patch series could be > a solution [8]. > > > Infrequent page list update problem > ==== > > Current page lists are updated by calling shrink_list() when memory pressure > comes, which might not be frequent enough to keep track of hot and cold pages. > Because all pages are on active lists at the first time shrink_list() is called > and the reference bit on the pages might not reflect the up to date access status > of these pages. But we also do not want to periodically shrink the global page > lists, which adds unnecessary overheads to the whole system. So I propose to > actively shrink page lists on the memcg we are interested in. > > Patch 18 to 25 add a new system call to shrink page lists on given application's > memcg and migrate pages between two NUMA nodes. It isolates the impact from the > rest of the system. To share DRAM among different applications, Patch 18 and 19 > add per-node memcg size limit, so you can limit the memory usage for particular > NUMA node(s). This sounds a little bit confusing to me. Is it totally user's decision about when to call the syscall to shrink page lists? But, how would user know when is a good timing? Could you please elaborate the usecase? Thanks, Yang > > > Patch structure > ==== > 1. multi-threaded page migration: Patch 01 to 06. > 2. DMA-based (using Intel IOAT DMA) page migration: Patch 07 and 08. > 3. concurrent (batched) page migration: Patch 09, 10, and 11. > 4. exchange pages: Patch 12 to 17. (This is a repost of part of [7]) > 5. per-node size limit in memcg: Patch 18 and 19. > 6. actively shrink page lists and perform page migration in given memcg: Patch 20 to 25. > > > Any comment is welcome. > > [1]: https://lore.kernel.org/linux-mm/20181226131446.330864849@intel.com/ > [2]: https://lore.kernel.org/linux-mm/20190321200157.29678-1-keith.busch@intel.com/ > [3]: https://lore.kernel.org/linux-mm/1553316275-21985-1-git-send-email-yang.shi@linux.alibaba.com/ > [4]: https://lore.kernel.org/linux-mm/6A903D34-A293-4056-B135-6FA227DE1828@nvidia.com/ > [5]: https://www.storagereview.com/supermicro_superserver_with_intel_optane_dc_persistent_memory_first_look_review > [6]: https://www.ibm.com/thought-leadership/summit-supercomputer/ > [7]: https://lore.kernel.org/linux-mm/20190215220856.29749-1-zi.yan@sent.com/ > [8]: https://lore.kernel.org/linux-mm/20181105165558.11698-1-daniel.m.jordan@oracle.com/ > > Zi Yan (25): > mm: migrate: Change migrate_mode to support combination migration > modes. > mm: migrate: Add mode parameter to support future page copy routines. > mm: migrate: Add a multi-threaded page migration function. > mm: migrate: Add copy_page_multithread into migrate_pages. > mm: migrate: Add vm.accel_page_copy in sysfs to control page copy > acceleration. > mm: migrate: Make the number of copy threads adjustable via sysctl. > mm: migrate: Add copy_page_dma to use DMA Engine to copy pages. > mm: migrate: Add copy_page_dma into migrate_page_copy. > mm: migrate: Add copy_page_lists_dma_always to support copy a list of > pages. > mm: migrate: copy_page_lists_mt() to copy a page list using > multi-threads. > mm: migrate: Add concurrent page migration into move_pages syscall. > exchange pages: new page migration mechanism: exchange_pages() > exchange pages: add multi-threaded exchange pages. > exchange pages: concurrent exchange pages. > exchange pages: exchange anonymous page and file-backed page. > exchange page: Add THP exchange support. > exchange page: Add exchange_page() syscall. > memcg: Add per node memory usage&max stats in memcg. > mempolicy: add MPOL_F_MEMCG flag, enforcing memcg memory limit. > memory manage: Add memory manage syscall. > mm: move update_lru_sizes() to mm_inline.h for broader use. > memory manage: active/inactive page list manipulation in memcg. > memory manage: page migration based page manipulation between NUMA > nodes. > memory manage: limit migration batch size. > memory manage: use exchange pages to memory manage to improve > throughput. > > arch/x86/entry/syscalls/syscall_64.tbl | 2 + > fs/aio.c | 12 +- > fs/f2fs/data.c | 6 +- > fs/hugetlbfs/inode.c | 4 +- > fs/iomap.c | 4 +- > fs/ubifs/file.c | 4 +- > include/linux/cgroup-defs.h | 1 + > include/linux/exchange.h | 27 + > include/linux/highmem.h | 3 + > include/linux/ksm.h | 4 + > include/linux/memcontrol.h | 67 ++ > include/linux/migrate.h | 12 +- > include/linux/migrate_mode.h | 8 + > include/linux/mm_inline.h | 21 + > include/linux/sched/coredump.h | 1 + > include/linux/sched/sysctl.h | 3 + > include/linux/syscalls.h | 10 + > include/uapi/linux/mempolicy.h | 9 +- > kernel/sysctl.c | 47 + > mm/Makefile | 5 + > mm/balloon_compaction.c | 2 +- > mm/compaction.c | 22 +- > mm/copy_page.c | 708 +++++++++++++++ > mm/exchange.c | 1560 ++++++++++++++++++++++++++++++++ > mm/exchange_page.c | 228 +++++ > mm/internal.h | 113 +++ > mm/ksm.c | 35 + > mm/memcontrol.c | 80 ++ > mm/memory_manage.c | 649 +++++++++++++ > mm/mempolicy.c | 38 +- > mm/migrate.c | 621 ++++++++++++- > mm/vmscan.c | 115 +-- > mm/zsmalloc.c | 2 +- > 33 files changed, 4261 insertions(+), 162 deletions(-) > create mode 100644 include/linux/exchange.h > create mode 100644 mm/copy_page.c > create mode 100644 mm/exchange.c > create mode 100644 mm/exchange_page.c > create mode 100644 mm/memory_manage.c > > -- > 2.7.4
>> Infrequent page list update problem >> ==== >> >> Current page lists are updated by calling shrink_list() when memory pressure >> comes, which might not be frequent enough to keep track of hot and cold pages. >> Because all pages are on active lists at the first time shrink_list() is called >> and the reference bit on the pages might not reflect the up to date access status >> of these pages. But we also do not want to periodically shrink the global page >> lists, which adds unnecessary overheads to the whole system. So I propose to >> actively shrink page lists on the memcg we are interested in. >> >> Patch 18 to 25 add a new system call to shrink page lists on given application's >> memcg and migrate pages between two NUMA nodes. It isolates the impact from the >> rest of the system. To share DRAM among different applications, Patch 18 and 19 >> add per-node memcg size limit, so you can limit the memory usage for particular >> NUMA node(s). > > This sounds a little bit confusing to me. Is it totally user's decision about when to call the syscall to shrink page lists? But, how would user know when is a good timing? Could you please elaborate the usecase? Sure. We would set up a daemon that monitors user applications and calls the syscall to shuffle the page lists for the user applications, although the daemon’s concrete action plan is still under exploration. It might not be ideal but the page access information could be refreshed periodically and page migration would happen on the background of application execution. On the other hand, if we wait until DRAM is full and use page migration to make room in DRAM for either page promotion or new page allocation, page migration sits on the critical path of application execution. Considering the bandwidth and access latency gaps between DRAM and PMEM are not as large as the gaps between DRAM and SSD, the cost of page migration (4KB/0.312GB/s = 12us or 2MB/2.387GB/s = 818us)might defeat the benefit of using DRAM over PMEM. I just wonder which would be better: waiting for 12us or 818us then reading 4KB or 2MB data in DRAM or directly accessing the data in PMEM without waiting. Let me know if this makes sense to you. Thanks. -- Best Regards, Yan Zi
From: Zi Yan <ziy@nvidia.com> Thanks to Dave Hansen's patches, which make PMEM as part of memory as NUMA nodes. How to use PMEM along with normal DRAM remains an open problem. There are several patchsets posted on the mailing list, proposing to use page migration to move pages between PMEM and DRAM using Linux page replacement policy [1,2,3]. There are some important problems not addressed in these patches: 1. The page migration in Linux does not provide high enough throughput for us to fully exploit PMEM or other use cases. 2. Linux page replacement is running too infrequent to distinguish hot and cold pages. I am trying to attack the problems with this patch series. This is not a final solution, but I would like to gather more feedback and comments from the mailing list. Page migration throughput problem ==== For example, in my recent email [4], I gave the page migration throughput numbers for different page migrations, none of which can achieve > 2.5GB/s throughput (the throughput is measured around kernel functions: migrate_pages() and migrate_page_copy()): | migrate_pages() | migrate_page_copy() migrating single 4KB page: | 0.312GB/s | 1.385GB/s migrating 512 4KB pages: | 0.854GB/s | 1.983GB/s migrating single 2MB THP: | 2.387GB/s | 2.481GB/s In reality, microbenchmarks show that Intel PMEM can provide ~65GB/s read throughput and ~16GB/s write throughput [5], which are much higher than the throughput achieved by Linux page migration. In addition, it is also desirable to use page migration to move data between high-bandwidth memory and DRAM, like IBM Summit, which exposes high-performance GPU memories as NUMA nodes [6]. This requires even higher page migration throughput. In this patch series, I propose four different ways of improving page migration throughput (mostly on 2MB THP migration): 1. multi-threaded page migration: Patch 03 to 06. 2. DMA-based (using Intel IOAT DMA) page migration: Patch 07 and 08. 3. concurrent (batched) page migration: Patch 09, 10, and 11. 4. exchange pages: Patch 12 to 17. (This is a repost of part of [7]) Here are some throughput numbers showing clear throughput improvements on a two-socket NUMA machine with two Xeon E5-2650 v3 @ 2.30GHz and a 19.2GB/s bandwidth QPI link (the same machine as mentioned in [4]): | migrate_pages() | migrate_page_copy() => migrating single 2MB THP | 2.387GB/s | 2.481GB/s 2-thread single THP migration | 3.478GB/s | 3.704GB/s 4-thread single THP migration | 5.474GB/s | 6.054GB/s 8-thread single THP migration | 7.846GB/s | 9.029GB/s 16-thread single THP migration | 7.423GB/s | 8.464GB/s 16-ch. DMA single THP migration | 4.322GB/s | 4.536GB/s 2-thread 16-THP migration | 3.610GB/s | 3.838GB/s 2-thread 16-THP batched migration | 4.138GB/s | 4.344GB/s 4-thread 16-THP migration | 6.385GB/s | 7.031GB/s 4-thread 16-THP batched migration | 7.382GB/s | 8.072GB/s 8-thread 16-THP migration | 8.039GB/s | 9.029GB/s 8-thread 16-THP batched migration | 9.023GB/s | 10.056GB/s 16-thread 16-THP migration | 8.137GB/s | 9.137GB/s 16-thread 16-THP batched migration | 9.907GB/s | 11.175GB/s 1-thread 16-THP exchange | 4.135GB/s | 4.225GB/s 2-thread 16-THP batched exchange | 7.061GB/s | 7.325GB/s 4-thread 16-THP batched exchange | 9.729GB/s | 10.237GB/s 8-thread 16-THP batched exchange | 9.992GB/s | 10.533GB/s 16-thread 16-THP batched exchange | 9.520GB/s | 10.056GB/s => migrating 512 4KB pages | 0.854GB/s | 1.983GB/s 1-thread 512-4KB batched exchange | 1.271GB/s | 3.433GB/s 2-thread 512-4KB batched exchange | 1.240GB/s | 3.190GB/s 4-thread 512-4KB batched exchange | 1.255GB/s | 3.823GB/s 8-thread 512-4KB batched exchange | 1.336GB/s | 3.921GB/s 16-thread 512-4KB batched exchange | 1.334GB/s | 3.897GB/s Concerns were raised on how to avoid CPU resource competition between page migration and user applications and have power awareness. Daniel Jordan recently posted a multi-threaded ktask patch series could be a solution [8]. Infrequent page list update problem ==== Current page lists are updated by calling shrink_list() when memory pressure comes, which might not be frequent enough to keep track of hot and cold pages. Because all pages are on active lists at the first time shrink_list() is called and the reference bit on the pages might not reflect the up to date access status of these pages. But we also do not want to periodically shrink the global page lists, which adds unnecessary overheads to the whole system. So I propose to actively shrink page lists on the memcg we are interested in. Patch 18 to 25 add a new system call to shrink page lists on given application's memcg and migrate pages between two NUMA nodes. It isolates the impact from the rest of the system. To share DRAM among different applications, Patch 18 and 19 add per-node memcg size limit, so you can limit the memory usage for particular NUMA node(s). Patch structure ==== 1. multi-threaded page migration: Patch 01 to 06. 2. DMA-based (using Intel IOAT DMA) page migration: Patch 07 and 08. 3. concurrent (batched) page migration: Patch 09, 10, and 11. 4. exchange pages: Patch 12 to 17. (This is a repost of part of [7]) 5. per-node size limit in memcg: Patch 18 and 19. 6. actively shrink page lists and perform page migration in given memcg: Patch 20 to 25. Any comment is welcome. [1]: https://lore.kernel.org/linux-mm/20181226131446.330864849@intel.com/ [2]: https://lore.kernel.org/linux-mm/20190321200157.29678-1-keith.busch@intel.com/ [3]: https://lore.kernel.org/linux-mm/1553316275-21985-1-git-send-email-yang.shi@linux.alibaba.com/ [4]: https://lore.kernel.org/linux-mm/6A903D34-A293-4056-B135-6FA227DE1828@nvidia.com/ [5]: https://www.storagereview.com/supermicro_superserver_with_intel_optane_dc_persistent_memory_first_look_review [6]: https://www.ibm.com/thought-leadership/summit-supercomputer/ [7]: https://lore.kernel.org/linux-mm/20190215220856.29749-1-zi.yan@sent.com/ [8]: https://lore.kernel.org/linux-mm/20181105165558.11698-1-daniel.m.jordan@oracle.com/ Zi Yan (25): mm: migrate: Change migrate_mode to support combination migration modes. mm: migrate: Add mode parameter to support future page copy routines. mm: migrate: Add a multi-threaded page migration function. mm: migrate: Add copy_page_multithread into migrate_pages. mm: migrate: Add vm.accel_page_copy in sysfs to control page copy acceleration. mm: migrate: Make the number of copy threads adjustable via sysctl. mm: migrate: Add copy_page_dma to use DMA Engine to copy pages. mm: migrate: Add copy_page_dma into migrate_page_copy. mm: migrate: Add copy_page_lists_dma_always to support copy a list of pages. mm: migrate: copy_page_lists_mt() to copy a page list using multi-threads. mm: migrate: Add concurrent page migration into move_pages syscall. exchange pages: new page migration mechanism: exchange_pages() exchange pages: add multi-threaded exchange pages. exchange pages: concurrent exchange pages. exchange pages: exchange anonymous page and file-backed page. exchange page: Add THP exchange support. exchange page: Add exchange_page() syscall. memcg: Add per node memory usage&max stats in memcg. mempolicy: add MPOL_F_MEMCG flag, enforcing memcg memory limit. memory manage: Add memory manage syscall. mm: move update_lru_sizes() to mm_inline.h for broader use. memory manage: active/inactive page list manipulation in memcg. memory manage: page migration based page manipulation between NUMA nodes. memory manage: limit migration batch size. memory manage: use exchange pages to memory manage to improve throughput. arch/x86/entry/syscalls/syscall_64.tbl | 2 + fs/aio.c | 12 +- fs/f2fs/data.c | 6 +- fs/hugetlbfs/inode.c | 4 +- fs/iomap.c | 4 +- fs/ubifs/file.c | 4 +- include/linux/cgroup-defs.h | 1 + include/linux/exchange.h | 27 + include/linux/highmem.h | 3 + include/linux/ksm.h | 4 + include/linux/memcontrol.h | 67 ++ include/linux/migrate.h | 12 +- include/linux/migrate_mode.h | 8 + include/linux/mm_inline.h | 21 + include/linux/sched/coredump.h | 1 + include/linux/sched/sysctl.h | 3 + include/linux/syscalls.h | 10 + include/uapi/linux/mempolicy.h | 9 +- kernel/sysctl.c | 47 + mm/Makefile | 5 + mm/balloon_compaction.c | 2 +- mm/compaction.c | 22 +- mm/copy_page.c | 708 +++++++++++++++ mm/exchange.c | 1560 ++++++++++++++++++++++++++++++++ mm/exchange_page.c | 228 +++++ mm/internal.h | 113 +++ mm/ksm.c | 35 + mm/memcontrol.c | 80 ++ mm/memory_manage.c | 649 +++++++++++++ mm/mempolicy.c | 38 +- mm/migrate.c | 621 ++++++++++++- mm/vmscan.c | 115 +-- mm/zsmalloc.c | 2 +- 33 files changed, 4261 insertions(+), 162 deletions(-) create mode 100644 include/linux/exchange.h create mode 100644 mm/copy_page.c create mode 100644 mm/exchange.c create mode 100644 mm/exchange_page.c create mode 100644 mm/memory_manage.c -- 2.7.4