Documentation/vm/ksm.txt

   1 How to use the Kernel Samepage Merging feature
   2 ----------------------------------------------
   3
   4 KSM is a memory-saving de-duplication feature, enabled by CONFIG_KSM=y,
   5 added to the Linux kernel in 2.6.32.  See mm/ksm.c for its implementation,
   6 and http://lwn.net/Articles/306704/ and http://lwn.net/Articles/330589/
   7
   8 The KSM daemon ksmd periodically scans those areas of user memory which
   9 have been registered with it, looking for pages of identical content which
  10 can be replaced by a single write-protected page (which is automatically
  11 copied if a process later wants to update its content).
  12
  13 KSM was originally developed for use with KVM (where it was known as
  14 Kernel Shared Memory), to fit more virtual machines into physical memory,
  15 by sharing the data common between them.  But it can be useful to any
  16 application which generates many instances of the same data.
  17
  18 KSM only merges anonymous (private) pages, never pagecache (file) pages.
  19 KSM's merged pages were originally locked into kernel memory, but can now
  20 be swapped out just like other user pages (but sharing is broken when they
  21 are swapped back in: ksmd must rediscover their identity and merge again).
  22
  23 KSM only operates on those areas of address space which an application
  24 has advised to be likely candidates for merging, by using the madvise(2)
  25 system call: int madvise(addr, length, MADV_MERGEABLE).
  26
  27 The app may call int madvise(addr, length, MADV_UNMERGEABLE) to cancel
  28 that advice and restore unshared pages: whereupon KSM unmerges whatever
  29 it merged in that range.  Note: this unmerging call may suddenly require
  30 more memory than is available - possibly failing with EAGAIN, but more
  31 probably arousing the Out-Of-Memory killer.
  32
  33 If KSM is not configured into the running kernel, madvise MADV_MERGEABLE
  34 and MADV_UNMERGEABLE simply fail with EINVAL.  If the running kernel was
  35 built with CONFIG_KSM=y, those calls will normally succeed: even if the
  36 the KSM daemon is not currently running, MADV_MERGEABLE still registers
  37 the range for whenever the KSM daemon is started; even if the range
  38 cannot contain any pages which KSM could actually merge; even if
  39 MADV_UNMERGEABLE is applied to a range which was never MADV_MERGEABLE.
  40
  41 Like other madvise calls, they are intended for use on mapped areas of
  42 the user address space: they will report ENOMEM if the specified range
  43 includes unmapped gaps (though working on the intervening mapped areas),
  44 and might fail with EAGAIN if not enough memory for internal structures.
  45
  46 Applications should be considerate in their use of MADV_MERGEABLE,
  47 restricting its use to areas likely to benefit.  KSM's scans may use a lot
  48 of processing power: some installations will disable KSM for that reason.
  49
  50 The KSM daemon is controlled by sysfs files in /sys/kernel/mm/ksm/,
  51 readable by all but writable only by root:
  52
  53 pages_to_scan    - how many present pages to scan before ksmd goes to sleep
  54                    e.g. "echo 100 > /sys/kernel/mm/ksm/pages_to_scan"
  55                    Default: 100 (chosen for demonstration purposes)
  56
  57 sleep_millisecs  - how many milliseconds ksmd should sleep before next scan
  58                    e.g. "echo 20 > /sys/kernel/mm/ksm/sleep_millisecs"
  59                    Default: 20 (chosen for demonstration purposes)
  60
  61 merge_across_nodes - specifies if pages from different numa nodes can be merged.
  62                    When set to 0, ksm merges only pages which physically
  63                    reside in the memory area of same NUMA node. That brings
  64                    lower latency to access of shared pages. Systems with more
  65                    nodes, at significant NUMA distances, are likely to benefit
  66                    from the lower latency of setting 0. Smaller systems, which
  67                    need to minimize memory usage, are likely to benefit from
  68                    the greater sharing of setting 1 (default). You may wish to
  69                    compare how your system performs under each setting, before
  70                    deciding on which to use. merge_across_nodes setting can be
  71                    changed only when there are no ksm shared pages in system:
  72                    set run 2 to unmerge pages first, then to 1 after changing
  73                    merge_across_nodes, to remerge according to the new setting.
  74                    Default: 1 (merging across nodes as in earlier releases)
  75
  76 run              - set 0 to stop ksmd from running but keep merged pages,
  77                    set 1 to run ksmd e.g. "echo 1 > /sys/kernel/mm/ksm/run",
  78                    set 2 to stop ksmd and unmerge all pages currently merged,
  79                          but leave mergeable areas registered for next run
  80                    Default: 0 (must be changed to 1 to activate KSM,
  81                                except if CONFIG_SYSFS is disabled)
  82
  83 use_zero_pages   - specifies whether empty pages (i.e. allocated pages
  84                    that only contain zeroes) should be treated specially.
  85                    When set to 1, empty pages are merged with the kernel
  86                    zero page(s) instead of with each other as it would
  87                    happen normally. This can improve the performance on
  88                    architectures with coloured zero pages, depending on
  89                    the workload. Care should be taken when enabling this
  90                    setting, as it can potentially degrade the performance
  91                    of KSM for some workloads, for example if the checksums
  92                    of pages candidate for merging match the checksum of
  93                    an empty page. This setting can be changed at any time,
  94                    it is only effective for pages merged after the change.
  95                    Default: 0 (normal KSM behaviour as in earlier releases)
  96
  97 max_page_sharing - Maximum sharing allowed for each KSM page. This
  98                    enforces a deduplication limit to avoid the virtual
  99                    memory rmap lists to grow too large. The minimum
 100                    value is 2 as a newly created KSM page will have at
 101                    least two sharers. The rmap walk has O(N)
 102                    complexity where N is the number of rmap_items
 103                    (i.e. virtual mappings) that are sharing the page,
 104                    which is in turn capped by max_page_sharing. So
 105                    this effectively spread the the linear O(N)
 106                    computational complexity from rmap walk context
 107                    over different KSM pages. The ksmd walk over the
 108                    stable_node "chains" is also O(N), but N is the
 109                    number of stable_node "dups", not the number of
 110                    rmap_items, so it has not a significant impact on
 111                    ksmd performance. In practice the best stable_node
 112                    "dup" candidate will be kept and found at the head
 113                    of the "dups" list. The higher this value the
 114                    faster KSM will merge the memory (because there
 115                    will be fewer stable_node dups queued into the
 116                    stable_node chain->hlist to check for pruning) and
 117                    the higher the deduplication factor will be, but
 118                    the slowest the worst case rmap walk could be for
 119                    any given KSM page. Slowing down the rmap_walk
 120                    means there will be higher latency for certain
 121                    virtual memory operations happening during
 122                    swapping, compaction, NUMA balancing and page
 123                    migration, in turn decreasing responsiveness for
 124                    the caller of those virtual memory operations. The
 125                    scheduler latency of other tasks not involved with
 126                    the VM operations doing the rmap walk is not
 127                    affected by this parameter as the rmap walks are
 128                    always schedule friendly themselves.
 129
 130 stable_node_chains_prune_millisecs - How frequently to walk the whole
 131                    list of stable_node "dups" linked in the
 132                    stable_node "chains" in order to prune stale
 133                    stable_nodes. Smaller milllisecs values will free
 134                    up the KSM metadata with lower latency, but they
 135                    will make ksmd use more CPU during the scan. This
 136                    only applies to the stable_node chains so it's a
 137                    noop if not a single KSM page hit the
 138                    max_page_sharing yet (there would be no stable_node
 139                    chains in such case).
 140
 141 The effectiveness of KSM and MADV_MERGEABLE is shown in /sys/kernel/mm/ksm/:
 142
 143 pages_shared     - how many shared pages are being used
 144 pages_sharing    - how many more sites are sharing them i.e. how much saved
 145 pages_unshared   - how many pages unique but repeatedly checked for merging
 146 pages_volatile   - how many pages changing too fast to be placed in a tree
 147 full_scans       - how many times all mergeable areas have been scanned
 148
 149 stable_node_chains - number of stable node chains allocated, this is
 150                      effectively the number of KSM pages that hit the
 151                      max_page_sharing limit
 152 stable_node_dups   - number of stable node dups queued into the
 153                      stable_node chains
 154
 155 A high ratio of pages_sharing to pages_shared indicates good sharing, but
 156 a high ratio of pages_unshared to pages_sharing indicates wasted effort.
 157 pages_volatile embraces several different kinds of activity, but a high
 158 proportion there would also indicate poor use of madvise MADV_MERGEABLE.
 159
 160 The maximum possible page_sharing/page_shared ratio is limited by the
 161 max_page_sharing tunable. To increase the ratio max_page_sharing must
 162 be increased accordingly.
 163
 164 The stable_node_dups/stable_node_chains ratio is also affected by the
 165 max_page_sharing tunable, and an high ratio may indicate fragmentation
 166 in the stable_node dups, which could be solved by introducing
 167 fragmentation algorithms in ksmd which would refile rmap_items from
 168 one stable_node dup to another stable_node dup, in order to freeup
 169 stable_node "dups" with few rmap_items in them, but that may increase
 170 the ksmd CPU usage and possibly slowdown the readonly computations on
 171 the KSM pages of the applications.
 172
 173 Izik Eidus,
 174 Hugh Dickins, 17 Nov 2009