3 =======================
4 Kernel Samepage Merging
5 =======================
10 KSM is a memory-saving de-duplication feature, enabled by CONFIG_KSM=y,
11 added to the Linux kernel in 2.6.32. See ``mm/ksm.c`` for its implementation,
12 and http://lwn.net/Articles/306704/ and https://lwn.net/Articles/330589/
14 KSM was originally developed for use with KVM (where it was known as
15 Kernel Shared Memory), to fit more virtual machines into physical memory,
16 by sharing the data common between them. But it can be useful to any
17 application which generates many instances of the same data.
19 The KSM daemon ksmd periodically scans those areas of user memory
20 which have been registered with it, looking for pages of identical
21 content which can be replaced by a single write-protected page (which
22 is automatically copied if a process later wants to update its
23 content). The amount of pages that KSM daemon scans in a single pass
24 and the time between the passes are configured using :ref:`sysfs
25 intraface <ksm_sysfs>`
27 KSM only merges anonymous (private) pages, never pagecache (file) pages.
28 KSM's merged pages were originally locked into kernel memory, but can now
29 be swapped out just like other user pages (but sharing is broken when they
30 are swapped back in: ksmd must rediscover their identity and merge again).
32 Controlling KSM with madvise
33 ============================
35 KSM only operates on those areas of address space which an application
36 has advised to be likely candidates for merging, by using the madvise(2)
39 int madvise(addr, length, MADV_MERGEABLE)
45 int madvise(addr, length, MADV_UNMERGEABLE)
47 to cancel that advice and restore unshared pages: whereupon KSM
48 unmerges whatever it merged in that range. Note: this unmerging call
49 may suddenly require more memory than is available - possibly failing
50 with EAGAIN, but more probably arousing the Out-Of-Memory killer.
52 If KSM is not configured into the running kernel, madvise MADV_MERGEABLE
53 and MADV_UNMERGEABLE simply fail with EINVAL. If the running kernel was
54 built with CONFIG_KSM=y, those calls will normally succeed: even if the
55 KSM daemon is not currently running, MADV_MERGEABLE still registers
56 the range for whenever the KSM daemon is started; even if the range
57 cannot contain any pages which KSM could actually merge; even if
58 MADV_UNMERGEABLE is applied to a range which was never MADV_MERGEABLE.
60 If a region of memory must be split into at least one new MADV_MERGEABLE
61 or MADV_UNMERGEABLE region, the madvise may return ENOMEM if the process
62 will exceed ``vm.max_map_count`` (see Documentation/admin-guide/sysctl/vm.rst).
64 Like other madvise calls, they are intended for use on mapped areas of
65 the user address space: they will report ENOMEM if the specified range
66 includes unmapped gaps (though working on the intervening mapped areas),
67 and might fail with EAGAIN if not enough memory for internal structures.
69 Applications should be considerate in their use of MADV_MERGEABLE,
70 restricting its use to areas likely to benefit. KSM's scans may use a lot
71 of processing power: some installations will disable KSM for that reason.
75 KSM daemon sysfs interface
76 ==========================
78 The KSM daemon is controlled by sysfs files in ``/sys/kernel/mm/ksm/``,
79 readable by all but writable only by root:
82 how many pages to scan before ksmd goes to sleep
83 e.g. ``echo 100 > /sys/kernel/mm/ksm/pages_to_scan``.
85 Default: 100 (chosen for demonstration purposes)
88 how many milliseconds ksmd should sleep before next scan
89 e.g. ``echo 20 > /sys/kernel/mm/ksm/sleep_millisecs``
91 Default: 20 (chosen for demonstration purposes)
94 specifies if pages from different NUMA nodes can be merged.
95 When set to 0, ksm merges only pages which physically reside
96 in the memory area of same NUMA node. That brings lower
97 latency to access of shared pages. Systems with more nodes, at
98 significant NUMA distances, are likely to benefit from the
99 lower latency of setting 0. Smaller systems, which need to
100 minimize memory usage, are likely to benefit from the greater
101 sharing of setting 1 (default). You may wish to compare how
102 your system performs under each setting, before deciding on
103 which to use. ``merge_across_nodes`` setting can be changed only
104 when there are no ksm shared pages in the system: set run 2 to
105 unmerge pages first, then to 1 after changing
106 ``merge_across_nodes``, to remerge according to the new setting.
108 Default: 1 (merging across nodes as in earlier releases)
111 * set to 0 to stop ksmd from running but keep merged pages,
112 * set to 1 to run ksmd e.g. ``echo 1 > /sys/kernel/mm/ksm/run``,
113 * set to 2 to stop ksmd and unmerge all pages currently merged, but
114 leave mergeable areas registered for next run.
116 Default: 0 (must be changed to 1 to activate KSM, except if
117 CONFIG_SYSFS is disabled)
120 specifies whether empty pages (i.e. allocated pages that only
121 contain zeroes) should be treated specially. When set to 1,
122 empty pages are merged with the kernel zero page(s) instead of
123 with each other as it would happen normally. This can improve
124 the performance on architectures with coloured zero pages,
125 depending on the workload. Care should be taken when enabling
126 this setting, as it can potentially degrade the performance of
127 KSM for some workloads, for example if the checksums of pages
128 candidate for merging match the checksum of an empty
129 page. This setting can be changed at any time, it is only
130 effective for pages merged after the change.
132 Default: 0 (normal KSM behaviour as in earlier releases)
135 Maximum sharing allowed for each KSM page. This enforces a
136 deduplication limit to avoid high latency for virtual memory
137 operations that involve traversal of the virtual mappings that
138 share the KSM page. The minimum value is 2 as a newly created
139 KSM page will have at least two sharers. The higher this value
140 the faster KSM will merge the memory and the higher the
141 deduplication factor will be, but the slower the worst case
142 virtual mappings traversal could be for any given KSM
143 page. Slowing down this traversal means there will be higher
144 latency for certain virtual memory operations happening during
145 swapping, compaction, NUMA balancing and page migration, in
146 turn decreasing responsiveness for the caller of those virtual
147 memory operations. The scheduler latency of other tasks not
148 involved with the VM operations doing the virtual mappings
149 traversal is not affected by this parameter as these
150 traversals are always schedule friendly themselves.
152 stable_node_chains_prune_millisecs
153 specifies how frequently KSM checks the metadata of the pages
154 that hit the deduplication limit for stale information.
155 Smaller milllisecs values will free up the KSM metadata with
156 lower latency, but they will make ksmd use more CPU during the
157 scan. It's a noop if not a single KSM page hit the
158 ``max_page_sharing`` yet.
160 The effectiveness of KSM and MADV_MERGEABLE is shown in ``/sys/kernel/mm/ksm/``:
163 how many shared pages are being used
165 how many more sites are sharing them i.e. how much saved
167 how many pages unique but repeatedly checked for merging
169 how many pages changing too fast to be placed in a tree
171 how many times all mergeable areas have been scanned
173 the number of KSM pages that hit the ``max_page_sharing`` limit
175 number of duplicated KSM pages
177 A high ratio of ``pages_sharing`` to ``pages_shared`` indicates good
178 sharing, but a high ratio of ``pages_unshared`` to ``pages_sharing``
179 indicates wasted effort. ``pages_volatile`` embraces several
180 different kinds of activity, but a high proportion there would also
181 indicate poor use of madvise MADV_MERGEABLE.
183 The maximum possible ``pages_sharing/pages_shared`` ratio is limited by the
184 ``max_page_sharing`` tunable. To increase the ratio ``max_page_sharing`` must
185 be increased accordingly.
189 Hugh Dickins, 17 Nov 2009