1 =============================
2 Guidance for writing policies
3 =============================
4
5 Try to keep transactionality out of it. The c
6 avoid asking about anything that is migrating.
7 makes it easier to write the policies.
8
9 Mappings are loaded into the policy at constru
10
11 Every bio that is mapped by the target is refe
12 The policy can return a simple HIT or MISS or
13
14 Currently there's no way for the policy to iss
15 e.g. to start writing back dirty blocks that a
16 soon.
17
18 Because we map bios, rather than requests it's
19 to get fooled by many small bios. For this re
20 issues periodic ticks to the policy. It's sug
21 doesn't update states (eg, hit counts) for a b
22 for each tick. The core ticks by watching bio
23 trying to see when the io scheduler has let th
24
25
26 Overview of supplied cache replacement policie
27 ==============================================
28
29 multiqueue (mq)
30 ---------------
31
32 This policy is now an alias for smq (see below
33
34 The following tunables are accepted, but have
35
36 'sequential_threshold <#nr_sequential_
37 'random_threshold <#nr_random_ios>'
38 'read_promote_adjustment <value>'
39 'write_promote_adjustment <value>'
40 'discard_promote_adjustment <value>'
41
42 Stochastic multiqueue (smq)
43 ---------------------------
44
45 This policy is the default.
46
47 The stochastic multi-queue (smq) policy addres
48 with the multiqueue (mq) policy.
49
50 The smq policy (vs mq) offers the promise of l
51 improved performance and increased adaptabilit
52 workloads. smq also does not have any cumbers
53
54 Users may switch from "mq" to "smq" simply by
55 DM table that is using the cache target. Doin
56 mq policy's hints to be dropped. Also, perfor
57 degrade slightly until smq recalculates the or
58 that should be cached.
59
60 Memory usage
61 ^^^^^^^^^^^^
62
63 The mq policy used a lot of memory; 88 bytes p
64 bit machine.
65
66 smq uses 28bit indexes to implement its data s
67 pointers. It avoids storing an explicit hit c
68 has a 'hotspot' queue, rather than a pre-cache
69 the entries (each hotspot block covers a large
70 cache block).
71
72 All this means smq uses ~25bytes per cache blo
73 memory, but a substantial improvement nonethel
74
75 Level balancing
76 ^^^^^^^^^^^^^^^
77
78 mq placed entries in different levels of the m
79 based on their hit count (~ln(hit count)). Th
80 levels generally had the most entries, and the
81 few. Having unbalanced levels like this reduc
82 multiqueue.
83
84 smq does not maintain a hit count, instead it
85 the least recently used entry from the level a
86 ordering being a side effect of this stochasti
87 scheme we can decide how many entries occupy e
88 resulting in better promotion/demotion decisio
89
90 Adaptability:
91 The mq policy maintained a hit count for each
92 different block to get promoted to the cache i
93 exceed the lowest currently in the cache. Thi
94 long time for the cache to adapt between varyi
95
96 smq doesn't maintain hit counts, so a lot of t
97 away. In addition it tracks performance of th
98 is used to decide which blocks to promote. If
99 performing badly then it starts moving entries
100 levels. This lets it adapt to new IO patterns
101
102 Performance
103 ^^^^^^^^^^^
104
105 Testing smq shows substantially better perform
106
107 cleaner
108 -------
109
110 The cleaner writes back all dirty blocks in a
111
112 Examples
113 ========
114
115 The syntax for a table is::
116
117 cache <metadata dev> <cache dev> <orig
118 <#feature_args> [<feature arg>]*
119 <policy> <#policy_args> [<policy arg>]
120
121 The syntax to send a message using the dmsetup
122
123 dmsetup message <mapped device> 0 sequ
124 dmsetup message <mapped device> 0 rand
125
126 Using dmsetup::
127
128 dmsetup create blah --table "0 2684354
129 /dev/sdd 512 0 mq 4 sequential_thr
130 creates a 128GB large mapped device na
131 sequential threshold set to 1024 and t
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.