sched-design-CFS.txt 9.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245
  1. =============
  2. CFS Scheduler
  3. =============
  4. 1. OVERVIEW
  5. CFS stands for "Completely Fair Scheduler," and is the new "desktop" process
  6. scheduler implemented by Ingo Molnar and merged in Linux 2.6.23. It is the
  7. replacement for the previous vanilla scheduler's SCHED_OTHER interactivity
  8. code.
  9. 80% of CFS's design can be summed up in a single sentence: CFS basically models
  10. an "ideal, precise multi-tasking CPU" on real hardware.
  11. "Ideal multi-tasking CPU" is a (non-existent :-)) CPU that has 100% physical
  12. power and which can run each task at precise equal speed, in parallel, each at
  13. 1/nr_running speed. For example: if there are 2 tasks running, then it runs
  14. each at 50% physical power --- i.e., actually in parallel.
  15. On real hardware, we can run only a single task at once, so we have to
  16. introduce the concept of "virtual runtime." The virtual runtime of a task
  17. specifies when its next timeslice would start execution on the ideal
  18. multi-tasking CPU described above. In practice, the virtual runtime of a task
  19. is its actual runtime normalized to the total number of running tasks.
  20. 2. FEW IMPLEMENTATION DETAILS
  21. In CFS the virtual runtime is expressed and tracked via the per-task
  22. p->se.vruntime (nanosec-unit) value. This way, it's possible to accurately
  23. timestamp and measure the "expected CPU time" a task should have gotten.
  24. [ small detail: on "ideal" hardware, at any time all tasks would have the same
  25. p->se.vruntime value --- i.e., tasks would execute simultaneously and no task
  26. would ever get "out of balance" from the "ideal" share of CPU time. ]
  27. CFS's task picking logic is based on this p->se.vruntime value and it is thus
  28. very simple: it always tries to run the task with the smallest p->se.vruntime
  29. value (i.e., the task which executed least so far). CFS always tries to split
  30. up CPU time between runnable tasks as close to "ideal multitasking hardware" as
  31. possible.
  32. Most of the rest of CFS's design just falls out of this really simple concept,
  33. with a few add-on embellishments like nice levels, multiprocessing and various
  34. algorithm variants to recognize sleepers.
  35. 3. THE RBTREE
  36. CFS's design is quite radical: it does not use the old data structures for the
  37. runqueues, but it uses a time-ordered rbtree to build a "timeline" of future
  38. task execution, and thus has no "array switch" artifacts (by which both the
  39. previous vanilla scheduler and RSDL/SD are affected).
  40. CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic
  41. increasing value tracking the smallest vruntime among all tasks in the
  42. runqueue. The total amount of work done by the system is tracked using
  43. min_vruntime; that value is used to place newly activated entities on the left
  44. side of the tree as much as possible.
  45. The total number of running tasks in the runqueue is accounted through the
  46. rq->cfs.load value, which is the sum of the weights of the tasks queued on the
  47. runqueue.
  48. CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the
  49. p->se.vruntime key (there is a subtraction using rq->cfs.min_vruntime to
  50. account for possible wraparounds). CFS picks the "leftmost" task from this
  51. tree and sticks to it.
  52. As the system progresses forwards, the executed tasks are put into the tree
  53. more and more to the right --- slowly but surely giving a chance for every task
  54. to become the "leftmost task" and thus get on the CPU within a deterministic
  55. amount of time.
  56. Summing up, CFS works like this: it runs a task a bit, and when the task
  57. schedules (or a scheduler tick happens) the task's CPU usage is "accounted
  58. for": the (small) time it just spent using the physical CPU is added to
  59. p->se.vruntime. Once p->se.vruntime gets high enough so that another task
  60. becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a
  61. small amount of "granularity" distance relative to the leftmost task so that we
  62. do not over-schedule tasks and trash the cache), then the new leftmost task is
  63. picked and the current task is preempted.
  64. 4. SOME FEATURES OF CFS
  65. CFS uses nanosecond granularity accounting and does not rely on any jiffies or
  66. other HZ detail. Thus the CFS scheduler has no notion of "timeslices" in the
  67. way the previous scheduler had, and has no heuristics whatsoever. There is
  68. only one central tunable (you have to switch on CONFIG_SCHED_DEBUG):
  69. /proc/sys/kernel/sched_min_granularity_ns
  70. which can be used to tune the scheduler from "desktop" (i.e., low latencies) to
  71. "server" (i.e., good batching) workloads. It defaults to a setting suitable
  72. for desktop workloads. SCHED_BATCH is handled by the CFS scheduler module too.
  73. Due to its design, the CFS scheduler is not prone to any of the "attacks" that
  74. exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c,
  75. chew.c, ring-test.c, massive_intr.c all work fine and do not impact
  76. interactivity and produce the expected behavior.
  77. The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH
  78. than the previous vanilla scheduler: both types of workloads are isolated much
  79. more aggressively.
  80. SMP load-balancing has been reworked/sanitized: the runqueue-walking
  81. assumptions are gone from the load-balancing code now, and iterators of the
  82. scheduling modules are used. The balancing code got quite a bit simpler as a
  83. result.
  84. 5. Scheduling policies
  85. CFS implements three scheduling policies:
  86. - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling
  87. policy that is used for regular tasks.
  88. - SCHED_BATCH: Does not preempt nearly as often as regular tasks
  89. would, thereby allowing tasks to run longer and make better use of
  90. caches but at the cost of interactivity. This is well suited for
  91. batch jobs.
  92. - SCHED_IDLE: This is even weaker than nice 19, but its not a true
  93. idle timer scheduler in order to avoid to get into priority
  94. inversion problems which would deadlock the machine.
  95. SCHED_FIFO/_RR are implemented in sched_rt.c and are as specified by
  96. POSIX.
  97. The command chrt from util-linux-ng 2.13.1.1 can set all of these except
  98. SCHED_IDLE.
  99. 6. SCHEDULING CLASSES
  100. The new CFS scheduler has been designed in such a way to introduce "Scheduling
  101. Classes," an extensible hierarchy of scheduler modules. These modules
  102. encapsulate scheduling policy details and are handled by the scheduler core
  103. without the core code assuming too much about them.
  104. sched_fair.c implements the CFS scheduler described above.
  105. sched_rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than
  106. the previous vanilla scheduler did. It uses 100 runqueues (for all 100 RT
  107. priority levels, instead of 140 in the previous scheduler) and it needs no
  108. expired array.
  109. Scheduling classes are implemented through the sched_class structure, which
  110. contains hooks to functions that must be called whenever an interesting event
  111. occurs.
  112. This is the (partial) list of the hooks:
  113. - enqueue_task(...)
  114. Called when a task enters a runnable state.
  115. It puts the scheduling entity (task) into the red-black tree and
  116. increments the nr_running variable.
  117. - dequeue_task(...)
  118. When a task is no longer runnable, this function is called to keep the
  119. corresponding scheduling entity out of the red-black tree. It decrements
  120. the nr_running variable.
  121. - yield_task(...)
  122. This function is basically just a dequeue followed by an enqueue, unless the
  123. compat_yield sysctl is turned on; in that case, it places the scheduling
  124. entity at the right-most end of the red-black tree.
  125. - check_preempt_curr(...)
  126. This function checks if a task that entered the runnable state should
  127. preempt the currently running task.
  128. - pick_next_task(...)
  129. This function chooses the most appropriate task eligible to run next.
  130. - set_curr_task(...)
  131. This function is called when a task changes its scheduling class or changes
  132. its task group.
  133. - task_tick(...)
  134. This function is mostly called from time tick functions; it might lead to
  135. process switch. This drives the running preemption.
  136. 7. GROUP SCHEDULER EXTENSIONS TO CFS
  137. Normally, the scheduler operates on individual tasks and strives to provide
  138. fair CPU time to each task. Sometimes, it may be desirable to group tasks and
  139. provide fair CPU time to each such task group. For example, it may be
  140. desirable to first provide fair CPU time to each user on the system and then to
  141. each task belonging to a user.
  142. CONFIG_CGROUP_SCHED strives to achieve exactly that. It lets tasks to be
  143. grouped and divides CPU time fairly among such groups.
  144. CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and
  145. SCHED_RR) tasks.
  146. CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and
  147. SCHED_BATCH) tasks.
  148. These options need CONFIG_CGROUPS to be defined, and let the administrator
  149. create arbitrary groups of tasks, using the "cgroup" pseudo filesystem. See
  150. Documentation/cgroups/cgroups.txt for more information about this filesystem.
  151. When CONFIG_FAIR_GROUP_SCHED is defined, a "cpu.shares" file is created for each
  152. group created using the pseudo filesystem. See example steps below to create
  153. task groups and modify their CPU share using the "cgroups" pseudo filesystem.
  154. # mount -t tmpfs cgroup_root /sys/fs/cgroup
  155. # mkdir /sys/fs/cgroup/cpu
  156. # mount -t cgroup -ocpu none /sys/fs/cgroup/cpu
  157. # cd /sys/fs/cgroup/cpu
  158. # mkdir multimedia # create "multimedia" group of tasks
  159. # mkdir browser # create "browser" group of tasks
  160. # #Configure the multimedia group to receive twice the CPU bandwidth
  161. # #that of browser group
  162. # echo 2048 > multimedia/cpu.shares
  163. # echo 1024 > browser/cpu.shares
  164. # firefox & # Launch firefox and move it to "browser" group
  165. # echo <firefox_pid> > browser/tasks
  166. # #Launch gmplayer (or your favourite movie player)
  167. # echo <movie_player_pid> > multimedia/tasks