as-iosched.txt 8.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167
  1. Anticipatory IO scheduler
  2. -------------------------
  3. Nick Piggin <piggin@cyberone.com.au> 13 Sep 2003
  4. Attention! Database servers, especially those using "TCQ" disks should
  5. investigate performance with the 'deadline' IO scheduler. Any system with high
  6. disk performance requirements should do so, in fact.
  7. If you see unusual performance characteristics of your disk systems, or you
  8. see big performance regressions versus the deadline scheduler, please email
  9. me. Database users don't bother unless you're willing to test a lot of patches
  10. from me ;) its a known issue.
  11. Also, users with hardware RAID controllers, doing striping, may find
  12. highly variable performance results with using the as-iosched. The
  13. as-iosched anticipatory implementation is based on the notion that a disk
  14. device has only one physical seeking head. A striped RAID controller
  15. actually has a head for each physical device in the logical RAID device.
  16. However, setting the antic_expire (see tunable parameters below) produces
  17. very similar behavior to the deadline IO scheduler.
  18. Selecting IO schedulers
  19. -----------------------
  20. To choose IO schedulers at boot time, use the argument 'elevator=deadline'.
  21. 'noop', 'as' and 'cfq' (the default) are also available. IO schedulers are
  22. assigned globally at boot time only presently. It's also possible to change
  23. the IO scheduler for a determined device on the fly, as described in
  24. Documentation/block/switching-sched.txt.
  25. Anticipatory IO scheduler Policies
  26. ----------------------------------
  27. The as-iosched implementation implements several layers of policies
  28. to determine when an IO request is dispatched to the disk controller.
  29. Here are the policies outlined, in order of application.
  30. 1. one-way Elevator algorithm.
  31. The elevator algorithm is similar to that used in deadline scheduler, with
  32. the addition that it allows limited backward movement of the elevator
  33. (i.e. seeks backwards). A seek backwards can occur when choosing between
  34. two IO requests where one is behind the elevator's current position, and
  35. the other is in front of the elevator's position. If the seek distance to
  36. the request in back of the elevator is less than half the seek distance to
  37. the request in front of the elevator, then the request in back can be chosen.
  38. Backward seeks are also limited to a maximum of MAXBACK (1024*1024) sectors.
  39. This favors forward movement of the elevator, while allowing opportunistic
  40. "short" backward seeks.
  41. 2. FIFO expiration times for reads and for writes.
  42. This is again very similar to the deadline IO scheduler. The expiration
  43. times for requests on these lists is tunable using the parameters read_expire
  44. and write_expire discussed below. When a read or a write expires in this way,
  45. the IO scheduler will interrupt its current elevator sweep or read anticipation
  46. to service the expired request.
  47. 3. Read and write request batching
  48. A batch is a collection of read requests or a collection of write
  49. requests. The as scheduler alternates dispatching read and write batches
  50. to the driver. In the case a read batch, the scheduler submits read
  51. requests to the driver as long as there are read requests to submit, and
  52. the read batch time limit has not been exceeded (read_batch_expire).
  53. The read batch time limit begins counting down only when there are
  54. competing write requests pending.
  55. In the case of a write batch, the scheduler submits write requests to
  56. the driver as long as there are write requests available, and the
  57. write batch time limit has not been exceeded (write_batch_expire).
  58. However, the length of write batches will be gradually shortened
  59. when read batches frequently exceed their time limit.
  60. When changing between batch types, the scheduler waits for all requests
  61. from the previous batch to complete before scheduling requests for the
  62. next batch.
  63. The read and write fifo expiration times described in policy 2 above
  64. are checked only when in scheduling IO of a batch for the corresponding
  65. (read/write) type. So for example, the read FIFO timeout values are
  66. tested only during read batches. Likewise, the write FIFO timeout
  67. values are tested only during write batches. For this reason,
  68. it is generally not recommended for the read batch time
  69. to be longer than the write expiration time, nor for the write batch
  70. time to exceed the read expiration time (see tunable parameters below).
  71. When the IO scheduler changes from a read to a write batch,
  72. it begins the elevator from the request that is on the head of the
  73. write expiration FIFO. Likewise, when changing from a write batch to
  74. a read batch, scheduler begins the elevator from the first entry
  75. on the read expiration FIFO.
  76. 4. Read anticipation.
  77. Read anticipation occurs only when scheduling a read batch.
  78. This implementation of read anticipation allows only one read request
  79. to be dispatched to the disk controller at a time. In
  80. contrast, many write requests may be dispatched to the disk controller
  81. at a time during a write batch. It is this characteristic that can make
  82. the anticipatory scheduler perform anomalously with controllers supporting
  83. TCQ, or with hardware striped RAID devices. Setting the antic_expire
  84. queue parameter (see below) to zero disables this behavior, and the
  85. anticipatory scheduler behaves essentially like the deadline scheduler.
  86. When read anticipation is enabled (antic_expire is not zero), reads
  87. are dispatched to the disk controller one at a time.
  88. At the end of each read request, the IO scheduler examines its next
  89. candidate read request from its sorted read list. If that next request
  90. is from the same process as the request that just completed,
  91. or if the next request in the queue is "very close" to the
  92. just completed request, it is dispatched immediately. Otherwise,
  93. statistics (average think time, average seek distance) on the process
  94. that submitted the just completed request are examined. If it seems
  95. likely that that process will submit another request soon, and that
  96. request is likely to be near the just completed request, then the IO
  97. scheduler will stop dispatching more read requests for up time (antic_expire)
  98. milliseconds, hoping that process will submit a new request near the one
  99. that just completed. If such a request is made, then it is dispatched
  100. immediately. If the antic_expire wait time expires, then the IO scheduler
  101. will dispatch the next read request from the sorted read queue.
  102. To decide whether an anticipatory wait is worthwhile, the scheduler
  103. maintains statistics for each process that can be used to compute
  104. mean "think time" (the time between read requests), and mean seek
  105. distance for that process. One observation is that these statistics
  106. are associated with each process, but those statistics are not associated
  107. with a specific IO device. So for example, if a process is doing IO
  108. on several file systems on separate devices, the statistics will be
  109. a combination of IO behavior from all those devices.
  110. Tuning the anticipatory IO scheduler
  111. ------------------------------------
  112. When using 'as', the anticipatory IO scheduler there are 5 parameters under
  113. /sys/block/*/queue/iosched/. All are units of milliseconds.
  114. The parameters are:
  115. * read_expire
  116. Controls how long until a read request becomes "expired". It also controls the
  117. interval between which expired requests are served, so set to 50, a request
  118. might take anywhere < 100ms to be serviced _if_ it is the next on the
  119. expired list. Obviously request expiration strategies won't make the disk
  120. go faster. The result basically equates to the timeslice a single reader
  121. gets in the presence of other IO. 100*((seek time / read_expire) + 1) is
  122. very roughly the % streaming read efficiency your disk should get with
  123. multiple readers.
  124. * read_batch_expire
  125. Controls how much time a batch of reads is given before pending writes are
  126. served. A higher value is more efficient. This might be set below read_expire
  127. if writes are to be given higher priority than reads, but reads are to be
  128. as efficient as possible when there are no writes. Generally though, it
  129. should be some multiple of read_expire.
  130. * write_expire, and
  131. * write_batch_expire are equivalent to the above, for writes.
  132. * antic_expire
  133. Controls the maximum amount of time we can anticipate a good read (one
  134. with a short seek distance from the most recently completed request) before
  135. giving up. Many other factors may cause anticipation to be stopped early,
  136. or some processes will not be "anticipated" at all. Should be a bit higher
  137. for big seek time devices though not a linear correspondence - most
  138. processes have only a few ms thinktime.