TEXT   137

cache.txt

Guest on 25th July 2021 05:25:10 PM

  1. Introduction
  2. ============
  3.  
  4. dm-cache is a device mapper target written by Joe Thornber, Heinz
  5. Mauelshagen, and Mike Snitzer.
  6.  
  7. It aims to improve performance of a block device (eg, a spindle) by
  8. dynamically migrating some of its data to a faster, smaller device
  9. (eg, an SSD).
  10.  
  11. This device-mapper solution allows us to insert this caching at
  12. different levels of the dm stack, for instance above the data device for
  13. a thin-provisioning pool.  Caching solutions that are integrated more
  14. closely with the virtual memory system should give better performance.
  15.  
  16. The target reuses the metadata library used in the thin-provisioning
  17. library.
  18.  
  19. The decision as to what data to migrate and when is left to a plug-in
  20. policy module.  Several of these have been written as we experiment,
  21. and we hope other people will contribute others for specific io
  22. scenarios (eg. a vm image server).
  23.  
  24. Glossary
  25. ========
  26.  
  27.   Migration -  Movement of the primary copy of a logical block from one
  28.                device to the other.
  29.   Promotion -  Migration from slow device to fast device.
  30.   Demotion  -  Migration from fast device to slow device.
  31.  
  32. The origin device always contains a copy of the logical block, which
  33. may be out of date or kept in sync with the copy on the cache device
  34. (depending on policy).
  35.  
  36. Design
  37. ======
  38.  
  39. Sub-devices
  40. -----------
  41.  
  42. The target is constructed by passing three devices to it (along with
  43. other parameters detailed later):
  44.  
  45. 1. An origin device - the big, slow one.
  46.  
  47. 2. A cache device - the small, fast one.
  48.  
  49. 3. A small metadata device - records which blocks are in the cache,
  50.    which are dirty, and extra hints for use by the policy object.
  51.    This information could be put on the cache device, but having it
  52.    separate allows the volume manager to configure it differently,
  53.    e.g. as a mirror for extra robustness.  This metadata device may only
  54.    be used by a single cache device.
  55.  
  56. Fixed block size
  57. ----------------
  58.  
  59. The origin is divided up into blocks of a fixed size.  This block size
  60. is configurable when you first create the cache.  Typically we've been
  61. using block sizes of 256KB - 1024KB.  The block size must be between 64
  62. (32KB) and 2097152 (1GB) and a multiple of 64 (32KB).
  63.  
  64. Having a fixed block size simplifies the target a lot.  But it is
  65. something of a compromise.  For instance, a small part of a block may be
  66. getting hit a lot, yet the whole block will be promoted to the cache.
  67. So large block sizes are bad because they waste cache space.  And small
  68. block sizes are bad because they increase the amount of metadata (both
  69. in core and on disk).
  70.  
  71. Writeback/writethrough
  72. ----------------------
  73.  
  74. The cache has two modes, writeback and writethrough.
  75.  
  76. If writeback, the default, is selected then a write to a block that is
  77. cached will go only to the cache and the block will be marked dirty in
  78. the metadata.
  79.  
  80. If writethrough is selected then a write to a cached block will not
  81. complete until it has hit both the origin and cache devices.  Clean
  82. blocks should remain clean.
  83.  
  84. A simple cleaner policy is provided, which will clean (write back) all
  85. dirty blocks in a cache.  Useful for decommissioning a cache.
  86.  
  87. Migration throttling
  88. --------------------
  89.  
  90. Migrating data between the origin and cache device uses bandwidth.
  91. The user can set a throttle to prevent more than a certain amount of
  92. migration occurring at any one time.  Currently we're not taking any
  93. account of normal io traffic going to the devices.  More work needs
  94. doing here to avoid migrating during those peak io moments.
  95.  
  96. For the time being, a message "migration_threshold <#sectors>"
  97. can be used to set the maximum number of sectors being migrated,
  98. the default being 204800 sectors (or 100MB).
  99.  
  100. Updating on-disk metadata
  101. -------------------------
  102.  
  103. On-disk metadata is committed every time a REQ_SYNC or REQ_FUA bio is
  104. written.  If no such requests are made then commits will occur every
  105. second.  This means the cache behaves like a physical disk that has a
  106. write cache (the same is true of the thin-provisioning target).  If
  107. power is lost you may lose some recent writes.  The metadata should
  108. always be consistent in spite of any crash.
  109.  
  110. The 'dirty' state for a cache block changes far too frequently for us
  111. to keep updating it on the fly.  So we treat it as a hint.  In normal
  112. operation it will be written when the dm device is suspended.  If the
  113. system crashes all cache blocks will be assumed dirty when restarted.
  114.  
  115. Per-block policy hints
  116. ----------------------
  117.  
  118. Policy plug-ins can store a chunk of data per cache block.  It's up to
  119. the policy how big this chunk is, but it should be kept small.  Like the
  120. dirty flags this data is lost if there's a crash so a safe fallback
  121. value should always be possible.
  122.  
  123. For instance, the 'mq' policy, which is currently the default policy,
  124. uses this facility to store the hit count of the cache blocks.  If
  125. there's a crash this information will be lost, which means the cache
  126. may be less efficient until those hit counts are regenerated.
  127.  
  128. Policy hints affect performance, not correctness.
  129.  
  130. Policy messaging
  131. ----------------
  132.  
  133. Policies will have different tunables, specific to each one, so we
  134. need a generic way of getting and setting these.  Device-mapper
  135. messages are used.  Refer to cache-policies.txt.
  136.  
  137. Discard bitset resolution
  138. -------------------------
  139.  
  140. We can avoid copying data during migration if we know the block has
  141. been discarded.  A prime example of this is when mkfs discards the
  142. whole block device.  We store a bitset tracking the discard state of
  143. blocks.  However, we allow this bitset to have a different block size
  144. from the cache blocks.  This is because we need to track the discard
  145. state for all of the origin device (compare with the dirty bitset
  146. which is just for the smaller cache device).
  147.  
  148. Target interface
  149. ================
  150.  
  151. Constructor
  152. -----------
  153.  
  154.  cache <metadata dev> <cache dev> <origin dev> <block size>
  155.        <#feature args> [<feature arg>]*
  156.        <policy> <#policy args> [policy args]*
  157.  
  158.  metadata dev    : fast device holding the persistent metadata
  159.  cache dev       : fast device holding cached data blocks
  160.  origin dev      : slow device holding original data blocks
  161.  block size      : cache unit size in sectors
  162.  
  163.  #feature args   : number of feature arguments passed
  164.  feature args    : writethrough.  (The default is writeback.)
  165.  
  166.  policy          : the replacement policy to use
  167.  #policy args    : an even number of arguments corresponding to
  168.                    key/value pairs passed to the policy
  169.  policy args     : key/value pairs passed to the policy
  170.                    E.g. 'sequential_threshold 1024'
  171.                    See cache-policies.txt for details.
  172.  
  173. Optional feature arguments are:
  174.    writethrough  : write through caching that prohibits cache block
  175.                    content from being different from origin block content.
  176.                    Without this argument, the default behaviour is to write
  177.                    back cache block contents later for performance reasons,
  178.                    so they may differ from the corresponding origin blocks.
  179.  
  180. A policy called 'default' is always registered.  This is an alias for
  181. the policy we currently think is giving best all round performance.
  182.  
  183. As the default policy could vary between kernels, if you are relying on
  184. the characteristics of a specific policy, always request it by name.
  185.  
  186. Status
  187. ------
  188.  
  189. <#used metadata blocks>/<#total metadata blocks> <#read hits> <#read misses>
  190. <#write hits> <#write misses> <#demotions> <#promotions> <#blocks in cache>
  191. <#dirty> <#features> <features>* <#core args> <core args>* <#policy args>
  192. <policy args>*
  193.  
  194. #used metadata blocks    : Number of metadata blocks used
  195. #total metadata blocks   : Total number of metadata blocks
  196. #read hits               : Number of times a READ bio has been mapped
  197.                              to the cache
  198. #read misses             : Number of times a READ bio has been mapped
  199.                              to the origin
  200. #write hits              : Number of times a WRITE bio has been mapped
  201.                              to the cache
  202. #write misses            : Number of times a WRITE bio has been
  203.                              mapped to the origin
  204. #demotions               : Number of times a block has been removed
  205.                              from the cache
  206. #promotions              : Number of times a block has been moved to
  207.                              the cache
  208. #blocks in cache         : Number of blocks resident in the cache
  209. #dirty                   : Number of blocks in the cache that differ
  210.                              from the origin
  211. #feature args            : Number of feature args to follow
  212. feature args             : 'writethrough' (optional)
  213. #core args               : Number of core arguments (must be even)
  214. core args                : Key/value pairs for tuning the core
  215.                              e.g. migration_threshold
  216. #policy args             : Number of policy arguments to follow (must be even)
  217. policy args              : Key/value pairs
  218.                              e.g. 'sequential_threshold 1024
  219.  
  220. Messages
  221. --------
  222.  
  223. Policies will have different tunables, specific to each one, so we
  224. need a generic way of getting and setting these.  Device-mapper
  225. messages are used.  (A sysfs interface would also be possible.)
  226.  
  227. The message format is:
  228.  
  229.    <key> <value>
  230.  
  231. E.g.
  232.    dmsetup message my_cache 0 sequential_threshold 1024
  233.  
  234. Examples
  235. ========
  236.  
  237. The test suite can be found here:
  238.  
  239. https://github.com/jthornber/thinp-test-suite
  240.  
  241. dmsetup create my_cache --table '0 41943040 cache /dev/mapper/metadata \
  242.         /dev/mapper/ssd /dev/mapper/origin 512 1 writeback default 0'
  243. dmsetup create my_cache --table '0 41943040 cache /dev/mapper/metadata \
  244.         /dev/mapper/ssd /dev/mapper/origin 1024 1 writeback \
  245.         mq 4 sequential_threshold 1024 random_threshold 8'

Raw Paste


Login or Register to edit or fork this paste. It's free.