1 =======================
2 Kernel Samepage Merging
3 =======================
8 KSM is a memory-saving de-duplication feature, enabled by CONFIG_KSM=y,
9 added to the Linux kernel in 2.6.32. See ``mm/ksm.c`` for its implementation,
10 and http://lwn.net/Articles/306704/ and https://lwn.net/Articles/330589/
12 KSM was originally developed for use with KVM (where it was known as
13 Kernel Shared Memory), to fit more virtual machines into physical memory,
14 by sharing the data common between them. But it can be useful to any
15 application which generates many instances of the same data.
17 The KSM daemon ksmd periodically scans those areas of user memory
18 which have been registered with it, looking for pages of identical
19 content which can be replaced by a single write-protected page (which
20 is automatically copied if a process later wants to update its
21 content). The amount of pages that KSM daemon scans in a single pass
22 and the time between the passes are configured using :ref:`sysfs
23 interface <ksm_sysfs>`
25 KSM only merges anonymous (private) pages, never pagecache (file) pages.
26 KSM's merged pages were originally locked into kernel memory, but can now
27 be swapped out just like other user pages (but sharing is broken when they
28 are swapped back in: ksmd must rediscover their identity and merge again).
30 Controlling KSM with madvise
31 ============================
33 KSM only operates on those areas of address space which an application
34 has advised to be likely candidates for merging, by using the madvise(2)
37 int madvise(addr, length, MADV_MERGEABLE)
43 int madvise(addr, length, MADV_UNMERGEABLE)
45 to cancel that advice and restore unshared pages: whereupon KSM
46 unmerges whatever it merged in that range. Note: this unmerging call
47 may suddenly require more memory than is available - possibly failing
48 with EAGAIN, but more probably arousing the Out-Of-Memory killer.
50 If KSM is not configured into the running kernel, madvise MADV_MERGEABLE
51 and MADV_UNMERGEABLE simply fail with EINVAL. If the running kernel was
52 built with CONFIG_KSM=y, those calls will normally succeed: even if the
53 KSM daemon is not currently running, MADV_MERGEABLE still registers
54 the range for whenever the KSM daemon is started; even if the range
55 cannot contain any pages which KSM could actually merge; even if
56 MADV_UNMERGEABLE is applied to a range which was never MADV_MERGEABLE.
58 If a region of memory must be split into at least one new MADV_MERGEABLE
59 or MADV_UNMERGEABLE region, the madvise may return ENOMEM if the process
60 will exceed ``vm.max_map_count`` (see Documentation/admin-guide/sysctl/vm.rst).
62 Like other madvise calls, they are intended for use on mapped areas of
63 the user address space: they will report ENOMEM if the specified range
64 includes unmapped gaps (though working on the intervening mapped areas),
65 and might fail with EAGAIN if not enough memory for internal structures.
67 Applications should be considerate in their use of MADV_MERGEABLE,
68 restricting its use to areas likely to benefit. KSM's scans may use a lot
69 of processing power: some installations will disable KSM for that reason.
73 KSM daemon sysfs interface
74 ==========================
76 The KSM daemon is controlled by sysfs files in ``/sys/kernel/mm/ksm/``,
77 readable by all but writable only by root:
80 how many pages to scan before ksmd goes to sleep
81 e.g. ``echo 100 > /sys/kernel/mm/ksm/pages_to_scan``.
83 Default: 100 (chosen for demonstration purposes)
86 how many milliseconds ksmd should sleep before next scan
87 e.g. ``echo 20 > /sys/kernel/mm/ksm/sleep_millisecs``
89 Default: 20 (chosen for demonstration purposes)
92 specifies if pages from different NUMA nodes can be merged.
93 When set to 0, ksm merges only pages which physically reside
94 in the memory area of same NUMA node. That brings lower
95 latency to access of shared pages. Systems with more nodes, at
96 significant NUMA distances, are likely to benefit from the
97 lower latency of setting 0. Smaller systems, which need to
98 minimize memory usage, are likely to benefit from the greater
99 sharing of setting 1 (default). You may wish to compare how
100 your system performs under each setting, before deciding on
101 which to use. ``merge_across_nodes`` setting can be changed only
102 when there are no ksm shared pages in the system: set run 2 to
103 unmerge pages first, then to 1 after changing
104 ``merge_across_nodes``, to remerge according to the new setting.
106 Default: 1 (merging across nodes as in earlier releases)
109 * set to 0 to stop ksmd from running but keep merged pages,
110 * set to 1 to run ksmd e.g. ``echo 1 > /sys/kernel/mm/ksm/run``,
111 * set to 2 to stop ksmd and unmerge all pages currently merged, but
112 leave mergeable areas registered for next run.
114 Default: 0 (must be changed to 1 to activate KSM, except if
115 CONFIG_SYSFS is disabled)
118 specifies whether empty pages (i.e. allocated pages that only
119 contain zeroes) should be treated specially. When set to 1,
120 empty pages are merged with the kernel zero page(s) instead of
121 with each other as it would happen normally. This can improve
122 the performance on architectures with coloured zero pages,
123 depending on the workload. Care should be taken when enabling
124 this setting, as it can potentially degrade the performance of
125 KSM for some workloads, for example if the checksums of pages
126 candidate for merging match the checksum of an empty
127 page. This setting can be changed at any time, it is only
128 effective for pages merged after the change.
130 Default: 0 (normal KSM behaviour as in earlier releases)
133 Maximum sharing allowed for each KSM page. This enforces a
134 deduplication limit to avoid high latency for virtual memory
135 operations that involve traversal of the virtual mappings that
136 share the KSM page. The minimum value is 2 as a newly created
137 KSM page will have at least two sharers. The higher this value
138 the faster KSM will merge the memory and the higher the
139 deduplication factor will be, but the slower the worst case
140 virtual mappings traversal could be for any given KSM
141 page. Slowing down this traversal means there will be higher
142 latency for certain virtual memory operations happening during
143 swapping, compaction, NUMA balancing and page migration, in
144 turn decreasing responsiveness for the caller of those virtual
145 memory operations. The scheduler latency of other tasks not
146 involved with the VM operations doing the virtual mappings
147 traversal is not affected by this parameter as these
148 traversals are always schedule friendly themselves.
150 stable_node_chains_prune_millisecs
151 specifies how frequently KSM checks the metadata of the pages
152 that hit the deduplication limit for stale information.
153 Smaller milllisecs values will free up the KSM metadata with
154 lower latency, but they will make ksmd use more CPU during the
155 scan. It's a noop if not a single KSM page hit the
156 ``max_page_sharing`` yet.
159 Historically KSM checked every candidate page for each scan. It did
160 not take into account historic information. When smart scan is
161 enabled, pages that have previously not been de-duplicated get
162 skipped. How often these pages are skipped depends on how often
163 de-duplication has already been tried and failed. By default this
164 optimization is enabled. The ``pages_skipped`` metric shows how
165 effective the setting is.
167 The effectiveness of KSM and MADV_MERGEABLE is shown in ``/sys/kernel/mm/ksm/``:
170 how effective is KSM. The calculation is explained below.
172 how many pages are being scanned for ksm
174 how many shared pages are being used
176 how many more sites are sharing them i.e. how much saved
178 how many pages unique but repeatedly checked for merging
180 how many pages changing too fast to be placed in a tree
182 how many pages did the "smart" page scanning algorithm skip
184 how many times all mergeable areas have been scanned
186 the number of KSM pages that hit the ``max_page_sharing`` limit
188 number of duplicated KSM pages
190 how many zero pages that are still mapped into processes were mapped by
191 KSM when deduplicating.
193 When ``use_zero_pages`` is/was enabled, the sum of ``pages_sharing`` +
194 ``ksm_zero_pages`` represents the actual number of pages saved by KSM.
195 if ``use_zero_pages`` has never been enabled, ``ksm_zero_pages`` is 0.
197 A high ratio of ``pages_sharing`` to ``pages_shared`` indicates good
198 sharing, but a high ratio of ``pages_unshared`` to ``pages_sharing``
199 indicates wasted effort. ``pages_volatile`` embraces several
200 different kinds of activity, but a high proportion there would also
201 indicate poor use of madvise MADV_MERGEABLE.
203 The maximum possible ``pages_sharing/pages_shared`` ratio is limited by the
204 ``max_page_sharing`` tunable. To increase the ratio ``max_page_sharing`` must
205 be increased accordingly.
207 Monitoring KSM profit
208 =====================
210 KSM can save memory by merging identical pages, but also can consume
211 additional memory, because it needs to generate a number of rmap_items to
212 save each scanned page's brief rmap information. Some of these pages may
213 be merged, but some may not be abled to be merged after being checked
214 several times, which are unprofitable memory consumed.
216 1) How to determine whether KSM save memory or consume memory in system-wide
217 range? Here is a simple approximate calculation for reference::
219 general_profit =~ ksm_saved_pages * sizeof(page) - (all_rmap_items) *
222 where ksm_saved_pages equals to the sum of ``pages_sharing`` +
223 ``ksm_zero_pages`` of the system, and all_rmap_items can be easily
224 obtained by summing ``pages_sharing``, ``pages_shared``, ``pages_unshared``
225 and ``pages_volatile``.
227 2) The KSM profit inner a single process can be similarly obtained by the
228 following approximate calculation::
230 process_profit =~ ksm_saved_pages * sizeof(page) -
231 ksm_rmap_items * sizeof(rmap_item).
233 where ksm_saved_pages equals to the sum of ``ksm_merging_pages`` and
234 ``ksm_zero_pages``, both of which are shown under the directory
235 ``/proc/<pid>/ksm_stat``, and ksm_rmap_items is also shown in
236 ``/proc/<pid>/ksm_stat``. The process profit is also shown in
237 ``/proc/<pid>/ksm_stat`` as ksm_process_profit.
239 From the perspective of application, a high ratio of ``ksm_rmap_items`` to
240 ``ksm_merging_pages`` means a bad madvise-applied policy, so developers or
241 administrators have to rethink how to change madvise policy. Giving an example
242 for reference, a page's size is usually 4K, and the rmap_item's size is
243 separately 32B on 32-bit CPU architecture and 64B on 64-bit CPU architecture.
244 so if the ``ksm_rmap_items/ksm_merging_pages`` ratio exceeds 64 on 64-bit CPU
245 or exceeds 128 on 32-bit CPU, then the app's madvise policy should be dropped,
246 because the ksm profit is approximately zero or negative.
248 Monitoring KSM events
249 =====================
251 There are some counters in /proc/vmstat that may be used to monitor KSM events.
252 KSM might help save memory, it's a tradeoff by may suffering delay on KSM COW
253 or on swapping in copy. Those events could help users evaluate whether or how
254 to use KSM. For example, if cow_ksm increases too fast, user may decrease the
255 range of madvise(, , MADV_MERGEABLE).
258 is incremented every time a KSM page triggers copy on write (COW)
259 when users try to write to a KSM page, we have to make a copy.
262 is incremented every time a KSM page is copied when swapping in
263 note that KSM page might be copied when swapping in because do_swap_page()
264 cannot do all the locking needed to reconstitute a cross-anon_vma KSM page.
268 Hugh Dickins, 17 Nov 2009