On Wed, Jul 15, 2015 at 6:54 AM, Vladimir Davydov <vdavydov@xxxxxxxxxxxxx> wrote: > Hi, > > This patch set introduces a new user API for tracking user memory pages > that have not been used for a given period of time. The purpose of this > is to provide the userspace with the means of tracking a workload's > working set, i.e. the set of pages that are actively used by the > workload. Knowing the working set size can be useful for partitioning > the system more efficiently, e.g. by tuning memory cgroup limits > appropriately, or for job placement within a compute cluster. > > It is based on top of v4.2-rc1-mmotm-2015-07-06-16-25 > > ---- USE CASES ---- > > The unified cgroup hierarchy has memory.low and memory.high knobs, which > are defined as the low and high boundaries for the workload working set > size. However, the working set size of a workload may be unknown or > change in time. With this patch set, one can periodically estimate the > amount of memory unused by each cgroup and tune their memory.low and > memory.high parameters accordingly, therefore optimizing the overall > memory utilization. > > Another use case is balancing workloads within a compute cluster. > Knowing how much memory is not really used by a workload unit may help > take a more optimal decision when considering migrating the unit to > another node within the cluster. > > Also, as noted by Minchan, this would be useful for per-process reclaim > (https://lwn.net/Articles/545668/). With idle tracking, we could reclaim idle > pages only by smart user memory manager. > > ---- USER API ---- > > The user API consists of two new proc files: > > * /proc/kpageidle. This file implements a bitmap where each bit corresponds > to a page, indexed by PFN. When the bit is set, the corresponding page is > idle. A page is considered idle if it has not been accessed since it was > marked idle. To mark a page idle one should set the bit corresponding to the > page by writing to the file. A value written to the file is OR-ed with the > current bitmap value. Only user memory pages can be marked idle, for other > page types input is silently ignored. Writing to this file beyond max PFN > results in the ENXIO error. Only available when CONFIG_IDLE_PAGE_TRACKING is > set. > > This file can be used to estimate the amount of pages that are not > used by a particular workload as follows: > > 1. mark all pages of interest idle by setting corresponding bits in the > /proc/kpageidle bitmap > 2. wait until the workload accesses its working set > 3. read /proc/kpageidle and count the number of bits set > > * /proc/kpagecgroup. This file contains a 64-bit inode number of the > memory cgroup each page is charged to, indexed by PFN. Only available when > CONFIG_MEMCG is set. > > This file can be used to find all pages (including unmapped file > pages) accounted to a particular cgroup. Using /proc/kpageidle, one > can then estimate the cgroup working set size. > > For an example of using these files for estimating the amount of unused > memory pages per each memory cgroup, please see the script attached > below. > > ---- REASONING ---- > > The reason to introduce the new user API instead of using > /proc/PID/{clear_refs,smaps} is that the latter has two serious > drawbacks: > > - it does not count unmapped file pages > - it affects the reclaimer logic > > The new API attempts to overcome them both. For more details on how it > is achieved, please see the comment to patch 5. > > ---- CHANGE LOG ---- > > Changes in v8: > > - clear referenced/accessed bit in secondary ptes while accessing > /proc/kpageidle; this is required to estimate wss of KVM VMs (Andres) > - check the young flag when collapsing a huge page > - copy idle/young flags on page migration Both good catches, thanks! I think the remaining question here is performance. Have you conducted any studies where - there is a workload - a daemon is poking kpageidle every N seconds/minutes - what is the daemon cpu consumption? - what is the workload degradation if any? N candidates include 30 seconds, 1 minute, 2 minutes, 5 minutes.... Workload candidates include TPC, spec int memory intensive things like 429.mcf, stream (http://www.cs.virginia.edu/stream/ "sustainable memory bandwidth" vs floating point performance) I'm not asking for a research paper, but if, say, a 2 minute-period daemon introduces no degradation and adds up to a minute of cpu per hour, then we're golden. Andres > > Changes in v7: > > This iteration addresses Andres's comments to v6: > > - do not reuse page_referenced for clearing idle flag, introduce a > separate function instead; this way we won't issue expensive tlb > flushes on /proc/kpageidle read/write > - propagate young/idle flags from head to tail pages on thp split > - skip compound tail pages while reading/writing /proc/kpageidle > - cleanup page_referenced_one > > Changes in v6: > > - Split the patch introducing page_cgroup_ino helper to ease review. > - Rebase on top of v4.1-rc7-mmotm-2015-06-09-16-55 > > Changes in v5: > > - Fix possible race between kpageidle_clear_pte_refs() and > __page_set_anon_rmap() by checking that a page is on an LRU list > under zone->lru_lock (Minchan). > - Export idle flag via /proc/kpageflags (Minchan). > - Rebase on top of 4.1-rc3. > > Changes in v4: > > This iteration primarily addresses Minchan's comments to v3: > > - Implement /proc/kpageidle as a bitmap instead of using u64 per each page, > because there does not seem to be any future uses for the other 63 bits. > - Do not double-increase pra->referenced in page_referenced_one() if the page > was young and referenced recently. > - Remove the pointless (page_count == 0) check from kpageidle_get_page(). > - Rename kpageidle_clear_refs() to kpageidle_clear_pte_refs(). > - Improve comments to kpageidle-related functions. > - Rebase on top of 4.1-rc2. > > Note it does not address Minchan's concern of possible __page_set_anon_rmap vs > page_referenced race (see https://lkml.org/lkml/2015/5/3/220) since it is still > unclear if this race can really happen (see https://lkml.org/lkml/2015/5/4/160) > > Changes in v3: > > - Enable CONFIG_IDLE_PAGE_TRACKING for 32 bit. Since this feature > requires two extra page flags and there is no space for them on 32 > bit, page ext is used (thanks to Minchan Kim). > - Minor code cleanups and comments improved. > - Rebase on top of 4.1-rc1. > > Changes in v2: > > - The main difference from v1 is the API change. In v1 the user can > only set the idle flag for all pages at once, and for clearing the > Idle flag on pages accessed via page tables /proc/PID/clear_refs > should be used. > The main drawback of the v1 approach, as noted by Minchan, is that on > big machines setting the idle flag for each pages can result in CPU > bursts, which would be especially frustrating if the user only wanted > to estimate the amount of idle pages for a particular process or VMA. > With the new API a more fine-grained approach is possible: one can > read a process's /proc/PID/pagemap and set/check the Idle flag only > for those pages of the process's address space he or she is > interested in. > Another good point about the v2 API is that it is possible to limit > /proc/kpage* scanning rate when the user wants to estimate the total > number of idle pages, which is unachievable with the v1 approach. > - Make /proc/kpagecgroup return the ino of the closest online ancestor > in case the cgroup a page is charged to is offline. > - Fix /proc/PID/clear_refs not clearing Young page flag. > - Rebase on top of v4.0-rc6-mmotm-2015-04-01-14-54 > > v7: https://lkml.org/lkml/2015/7/11/119 > v6: https://lkml.org/lkml/2015/6/12/301 > v5: https://lkml.org/lkml/2015/5/12/449 > v4: https://lkml.org/lkml/2015/5/7/580 > v3: https://lkml.org/lkml/2015/4/28/224 > v2: https://lkml.org/lkml/2015/4/7/260 > v1: https://lkml.org/lkml/2015/3/18/794 > > ---- PATCH SET STRUCTURE ---- > > The patch set is organized as follows: > > - patch 1 adds page_cgroup_ino() helper for the sake of > /proc/kpagecgroup and patches 2-3 do related cleanup > - patch 4 adds /proc/kpagecgroup, which reports cgroup ino each page is > charged to > - patch 5 introduces a new mmu notifier callback, clear_young, which is > a lightweight version of clear_flush_young; it is used in patch 6 > - patch 6 implements the idle page tracking feature, including the > userspace API, /proc/kpageidle > - patch 7 exports idle flag via /proc/kpageflags > > ---- SIMILAR WORKS ---- > > Originally, the patch for tracking idle memory was proposed back in 2011 > by Michel Lespinasse (see http://lwn.net/Articles/459269/). The main > difference between Michel's patch and this one is that Michel > implemented a kernel space daemon for estimating idle memory size per > cgroup while this patch only provides the userspace with the minimal API > for doing the job, leaving the rest up to the userspace. However, they > both share the same idea of Idle/Young page flags to avoid affecting the > reclaimer logic. > > ---- SCRIPT FOR COUNTING IDLE PAGES PER CGROUP ---- > #! /usr/bin/python > # > > import os > import stat > import errno > import struct > > CGROUP_MOUNT = "/sys/fs/cgroup/memory" > BUFSIZE = 8 * 1024 # must be multiple of 8 > > > def get_hugepage_size(): > with open("/proc/meminfo", "r") as f: > for s in f: > k, v = s.split(":") > if k == "Hugepagesize": > return int(v.split()[0]) * 1024 > > PAGE_SIZE = os.sysconf("SC_PAGE_SIZE") > HUGEPAGE_SIZE = get_hugepage_size() > > > def set_idle(): > f = open("/proc/kpageidle", "wb", BUFSIZE) > while True: > try: > f.write(struct.pack("Q", pow(2, 64) - 1)) > except IOError as err: > if err.errno == errno.ENXIO: > break > raise > f.close() > > > def count_idle(): > f_flags = open("/proc/kpageflags", "rb", BUFSIZE) > f_cgroup = open("/proc/kpagecgroup", "rb", BUFSIZE) > > with open("/proc/kpageidle", "rb", BUFSIZE) as f: > while f.read(BUFSIZE): pass # update idle flag > > idlememsz = {} > while True: > s1, s2 = f_flags.read(8), f_cgroup.read(8) > if not s1 or not s2: > break > > flags, = struct.unpack('Q', s1) > cgino, = struct.unpack('Q', s2) > > unevictable = (flags >> 18) & 1 > huge = (flags >> 22) & 1 > idle = (flags >> 25) & 1 > > if idle and not unevictable: > idlememsz[cgino] = idlememsz.get(cgino, 0) + \ > (HUGEPAGE_SIZE if huge else PAGE_SIZE) > > f_flags.close() > f_cgroup.close() > return idlememsz > > > if __name__ == "__main__": > print "Setting the idle flag for each page..." > set_idle() > > raw_input("Wait until the workload accesses its working set, " > "then press Enter") > > print "Counting idle pages..." > idlememsz = count_idle() > > for dir, subdirs, files in os.walk(CGROUP_MOUNT): > ino = os.stat(dir)[stat.ST_INO] > print dir + ": " + str(idlememsz.get(ino, 0) / 1024) + " kB" > ---- END SCRIPT ---- > > Comments are more than welcome. > > Thanks, > > Vladimir Davydov (7): > memcg: add page_cgroup_ino helper > hwpoison: use page_cgroup_ino for filtering by memcg > memcg: zap try_get_mem_cgroup_from_page > proc: add kpagecgroup file > mmu-notifier: add clear_young callback > proc: add kpageidle file > proc: export idle flag via kpageflags > > Documentation/vm/pagemap.txt | 22 ++- > fs/proc/page.c | 274 +++++++++++++++++++++++++++++++++ > fs/proc/task_mmu.c | 4 +- > include/linux/memcontrol.h | 7 +- > include/linux/mm.h | 98 ++++++++++++ > include/linux/mmu_notifier.h | 44 ++++++ > include/linux/page-flags.h | 11 ++ > include/linux/page_ext.h | 4 + > include/uapi/linux/kernel-page-flags.h | 1 + > mm/Kconfig | 12 ++ > mm/debug.c | 4 + > mm/huge_memory.c | 11 +- > mm/hwpoison-inject.c | 5 +- > mm/memcontrol.c | 71 +++++---- > mm/memory-failure.c | 16 +- > mm/migrate.c | 5 + > mm/mmu_notifier.c | 17 ++ > mm/page_ext.c | 3 + > mm/rmap.c | 5 + > mm/swap.c | 2 + > virt/kvm/kvm_main.c | 18 +++ > 21 files changed, 570 insertions(+), 64 deletions(-) > > -- > 2.1.4 > -- Andres Lagar-Cavilla | Google Kernel Team | andreslc@xxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-doc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
