| blob | 6686bd267dc9d35a66a02fb4590aeece99b2171e |
1 How to use the Kernel Samepage Merging feature
2 ----------------------------------------------
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/
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).
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.
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).
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).
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.
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.
41 If a region of memory must be split into at least one new MADV_MERGEABLE
42 or MADV_UNMERGEABLE region, the madvise may return ENOMEM if the process
43 will exceed vm.max_map_count (see Documentation/sysctl/vm.txt).
45 Like other madvise calls, they are intended for use on mapped areas of
46 the user address space: they will report ENOMEM if the specified range
47 includes unmapped gaps (though working on the intervening mapped areas),
48 and might fail with EAGAIN if not enough memory for internal structures.
50 Applications should be considerate in their use of MADV_MERGEABLE,
51 restricting its use to areas likely to benefit. KSM's scans may use a lot
52 of processing power: some installations will disable KSM for that reason.
54 The KSM daemon is controlled by sysfs files in /sys/kernel/mm/ksm/,
55 readable by all but writable only by root:
57 pages_to_scan - how many present pages to scan before ksmd goes to sleep
58 e.g. "echo 100 > /sys/kernel/mm/ksm/pages_to_scan"
59 Default: 100 (chosen for demonstration purposes)
61 sleep_millisecs - how many milliseconds ksmd should sleep before next scan
62 e.g. "echo 20 > /sys/kernel/mm/ksm/sleep_millisecs"
63 Default: 20 (chosen for demonstration purposes)
65 merge_across_nodes - specifies if pages from different numa nodes can be merged.
66 When set to 0, ksm merges only pages which physically
67 reside in the memory area of same NUMA node. That brings
68 lower latency to access of shared pages. Systems with more
69 nodes, at significant NUMA distances, are likely to benefit
70 from the lower latency of setting 0. Smaller systems, which
71 need to minimize memory usage, are likely to benefit from
72 the greater sharing of setting 1 (default). You may wish to
73 compare how your system performs under each setting, before
74 deciding on which to use. merge_across_nodes setting can be
75 changed only when there are no ksm shared pages in system:
76 set run 2 to unmerge pages first, then to 1 after changing
77 merge_across_nodes, to remerge according to the new setting.
78 Default: 1 (merging across nodes as in earlier releases)
80 run - set 0 to stop ksmd from running but keep merged pages,
81 set 1 to run ksmd e.g. "echo 1 > /sys/kernel/mm/ksm/run",
82 set 2 to stop ksmd and unmerge all pages currently merged,
83 but leave mergeable areas registered for next run
84 Default: 0 (must be changed to 1 to activate KSM,
85 except if CONFIG_SYSFS is disabled)
87 use_zero_pages - specifies whether empty pages (i.e. allocated pages
88 that only contain zeroes) should be treated specially.
89 When set to 1, empty pages are merged with the kernel
90 zero page(s) instead of with each other as it would
91 happen normally. This can improve the performance on
92 architectures with coloured zero pages, depending on
93 the workload. Care should be taken when enabling this
94 setting, as it can potentially degrade the performance
95 of KSM for some workloads, for example if the checksums
96 of pages candidate for merging match the checksum of
97 an empty page. This setting can be changed at any time,
98 it is only effective for pages merged after the change.
99 Default: 0 (normal KSM behaviour as in earlier releases)
101 max_page_sharing - Maximum sharing allowed for each KSM page. This
102 enforces a deduplication limit to avoid the virtual
103 memory rmap lists to grow too large. The minimum
104 value is 2 as a newly created KSM page will have at
105 least two sharers. The rmap walk has O(N)
106 complexity where N is the number of rmap_items
107 (i.e. virtual mappings) that are sharing the page,
108 which is in turn capped by max_page_sharing. So
109 this effectively spread the the linear O(N)
110 computational complexity from rmap walk context
111 over different KSM pages. The ksmd walk over the
112 stable_node "chains" is also O(N), but N is the
113 number of stable_node "dups", not the number of
114 rmap_items, so it has not a significant impact on
115 ksmd performance. In practice the best stable_node
116 "dup" candidate will be kept and found at the head
117 of the "dups" list. The higher this value the
118 faster KSM will merge the memory (because there
119 will be fewer stable_node dups queued into the
120 stable_node chain->hlist to check for pruning) and
121 the higher the deduplication factor will be, but
122 the slowest the worst case rmap walk could be for
123 any given KSM page. Slowing down the rmap_walk
124 means there will be higher latency for certain
125 virtual memory operations happening during
126 swapping, compaction, NUMA balancing and page
127 migration, in turn decreasing responsiveness for
128 the caller of those virtual memory operations. The
129 scheduler latency of other tasks not involved with
130 the VM operations doing the rmap walk is not
131 affected by this parameter as the rmap walks are
132 always schedule friendly themselves.
134 stable_node_chains_prune_millisecs - How frequently to walk the whole
135 list of stable_node "dups" linked in the
136 stable_node "chains" in order to prune stale
137 stable_nodes. Smaller milllisecs values will free
138 up the KSM metadata with lower latency, but they
139 will make ksmd use more CPU during the scan. This
140 only applies to the stable_node chains so it's a
141 noop if not a single KSM page hit the
142 max_page_sharing yet (there would be no stable_node
143 chains in such case).
145 The effectiveness of KSM and MADV_MERGEABLE is shown in /sys/kernel/mm/ksm/:
147 pages_shared - how many shared pages are being used
148 pages_sharing - how many more sites are sharing them i.e. how much saved
149 pages_unshared - how many pages unique but repeatedly checked for merging
150 pages_volatile - how many pages changing too fast to be placed in a tree
151 full_scans - how many times all mergeable areas have been scanned
153 stable_node_chains - number of stable node chains allocated, this is
154 effectively the number of KSM pages that hit the
155 max_page_sharing limit
156 stable_node_dups - number of stable node dups queued into the
157 stable_node chains
159 A high ratio of pages_sharing to pages_shared indicates good sharing, but
160 a high ratio of pages_unshared to pages_sharing indicates wasted effort.
161 pages_volatile embraces several different kinds of activity, but a high
162 proportion there would also indicate poor use of madvise MADV_MERGEABLE.
164 The maximum possible page_sharing/page_shared ratio is limited by the
165 max_page_sharing tunable. To increase the ratio max_page_sharing must
166 be increased accordingly.
168 The stable_node_dups/stable_node_chains ratio is also affected by the
169 max_page_sharing tunable, and an high ratio may indicate fragmentation
170 in the stable_node dups, which could be solved by introducing
171 fragmentation algorithms in ksmd which would refile rmap_items from
172 one stable_node dup to another stable_node dup, in order to freeup
173 stable_node "dups" with few rmap_items in them, but that may increase
174 the ksmd CPU usage and possibly slowdown the readonly computations on
175 the KSM pages of the applications.
177 Izik Eidus,
178 Hugh Dickins, 17 Nov 2009