summaryrefslogtreecommitdiffstats
path: root/Documentation/device-mapper/cache-policies.txt
blob: d9246a32e673f273c03bdd2a0da9c44373c11891 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
Guidance for writing policies
=============================

Try to keep transactionality out of it.  The core is careful to
avoid asking about anything that is migrating.  This is a pain, but
makes it easier to write the policies.

Mappings are loaded into the policy at construction time.

Every bio that is mapped by the target is referred to the policy.
The policy can return a simple HIT or MISS or issue a migration.

Currently there's no way for the policy to issue background work,
e.g. to start writing back dirty blocks that are going to be evicte
soon.

Because we map bios, rather than requests it's easy for the policy
to get fooled by many small bios.  For this reason the core target
issues periodic ticks to the policy.  It's suggested that the policy
doesn't update states (eg, hit counts) for a block more than once
for each tick.  The core ticks by watching bios complete, and so
trying to see when the io scheduler has let the ios run.


Overview of supplied cache replacement policies
===============================================

multiqueue (mq)
---------------

This policy has been deprecated in favor of the smq policy (see below).

The multiqueue policy has three sets of 16 queues: one set for entries
waiting for the cache and another two for those in the cache (a set for
clean entries and a set for dirty entries).

Cache entries in the queues are aged based on logical time. Entry into
the cache is based on variable thresholds and queue selection is based
on hit count on entry. The policy aims to take different cache miss
costs into account and to adjust to varying load patterns automatically.

Message and constructor argument pairs are:
	'sequential_threshold <#nr_sequential_ios>'
	'random_threshold <#nr_random_ios>'
	'read_promote_adjustment <value>'
	'write_promote_adjustment <value>'
	'discard_promote_adjustment <value>'

The sequential threshold indicates the number of contiguous I/Os
required before a stream is treated as sequential.  Once a stream is
considered sequential it will bypass the cache.  The random threshold
is the number of intervening non-contiguous I/Os that must be seen
before the stream is treated as random again.

The sequential and random thresholds default to 512 and 4 respectively.

Large, sequential I/Os are probably better left on the origin device
since spindles tend to have good sequential I/O bandwidth.  The
io_tracker counts contiguous I/Os to try to spot when the I/O is in one
of these sequential modes.  But there are use-cases for wanting to
promote sequential blocks to the cache (e.g. fast application startup).
If sequential threshold is set to 0 the sequential I/O detection is
disabled and sequential I/O will no longer implicitly bypass the cache.
Setting the random threshold to 0 does _not_ disable the random I/O
stream detection.

Internally the mq policy determines a promotion threshold.  If the hit
count of a block not in the cache goes above this threshold it gets
promoted to the cache.  The read, write and discard promote adjustment
tunables allow you to tweak the promotion threshold by adding a small
value based on the io type.  They default to 4, 8 and 1 respectively.
If you're trying to quickly warm a new cache device you may wish to
reduce these to encourage promotion.  Remember to switch them back to
their defaults after the cache fills though.

Stochastic multiqueue (smq)
---------------------------

This policy is the default.

The stochastic multi-queue (smq) policy addresses some of the problems
with the multiqueue (mq) policy.

The smq policy (vs mq) offers the promise of less memory utilization,
improved performance and increased adaptability in the face of changing
workloads.  SMQ also does not have any cumbersome tuning knobs.

Users may switch from "mq" to "smq" simply by appropriately reloading a
DM table that is using the cache target.  Doing so will cause all of the
mq policy's hints to be dropped.  Also, performance of the cache may
degrade slightly until smq recalculates the origin device's hotspots
that should be cached.

Memory usage:
The mq policy uses a lot of memory; 88 bytes per cache block on a 64
bit machine.

SMQ uses 28bit indexes to implement it's data structures rather than
pointers.  It avoids storing an explicit hit count for each block.  It
has a 'hotspot' queue rather than a pre cache which uses a quarter of
the entries (each hotspot block covers a larger area than a single
cache block).

All these mean smq uses ~25bytes per cache block.  Still a lot of
memory, but a substantial improvement nontheless.

Level balancing:
MQ places entries in different levels of the multiqueue structures
based on their hit count (~ln(hit count)).  This means the bottom
levels generally have the most entries, and the top ones have very
few.  Having unbalanced levels like this reduces the efficacy of the
multiqueue.

SMQ does not maintain a hit count, instead it swaps hit entries with
the least recently used entry from the level above.  The over all
ordering being a side effect of this stochastic process.  With this
scheme we can decide how many entries occupy each multiqueue level,
resulting in better promotion/demotion decisions.

Adaptability:
The MQ policy maintains a hit count for each cache block.  For a
different block to get promoted to the cache it's hit count has to
exceed the lowest currently in the cache.  This means it can take a
long time for the cache to adapt between varying IO patterns.
Periodically degrading the hit counts could help with this, but I
haven't found a nice general solution.

SMQ doesn't maintain hit counts, so a lot of this problem just goes
away.  In addition it tracks performance of the hotspot queue, which
is used to decide which blocks to promote.  If the hotspot queue is
performing badly then it starts moving entries more quickly between
levels.  This lets it adapt to new IO patterns very quickly.

Performance:
Testing SMQ shows substantially better performance than MQ.

cleaner
-------

The cleaner writes back all dirty blocks in a cache to decommission it.

Examples
========

The syntax for a table is:
	cache <metadata dev> <cache dev> <origin dev> <block size>
	<#feature_args> [<feature arg>]*
	<policy> <#policy_args> [<policy arg>]*

The syntax to send a message using the dmsetup command is:
	dmsetup message <mapped device> 0 sequential_threshold 1024
	dmsetup message <mapped device> 0 random_threshold 8

Using dmsetup:
	dmsetup create blah --table "0 268435456 cache /dev/sdb /dev/sdc \
	    /dev/sdd 512 0 mq 4 sequential_threshold 1024 random_threshold 8"
	creates a 128GB large mapped device named 'blah' with the
	sequential threshold set to 1024 and the random_threshold set to 8.
OpenPOWER on IntegriCloud