sched-design-CFS.rst 9.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249
  1. =============
  2. CFS Scheduler
  3. =============
  4. 1. OVERVIEW
  5. ============
  6. CFS stands for "Completely Fair Scheduler," and is the new "desktop" process
  7. scheduler implemented by Ingo Molnar and merged in Linux 2.6.23. It is the
  8. replacement for the previous vanilla scheduler's SCHED_OTHER interactivity
  9. code.
  10. 80% of CFS's design can be summed up in a single sentence: CFS basically models
  11. an "ideal, precise multi-tasking CPU" on real hardware.
  12. "Ideal multi-tasking CPU" is a (non-existent :-)) CPU that has 100% physical
  13. power and which can run each task at precise equal speed, in parallel, each at
  14. 1/nr_running speed. For example: if there are 2 tasks running, then it runs
  15. each at 50% physical power --- i.e., actually in parallel.
  16. On real hardware, we can run only a single task at once, so we have to
  17. introduce the concept of "virtual runtime." The virtual runtime of a task
  18. specifies when its next timeslice would start execution on the ideal
  19. multi-tasking CPU described above. In practice, the virtual runtime of a task
  20. is its actual runtime normalized to the total number of running tasks.
  21. 2. FEW IMPLEMENTATION DETAILS
  22. ==============================
  23. In CFS the virtual runtime is expressed and tracked via the per-task
  24. p->se.vruntime (nanosec-unit) value. This way, it's possible to accurately
  25. timestamp and measure the "expected CPU time" a task should have gotten.
  26. [ small detail: on "ideal" hardware, at any time all tasks would have the same
  27. p->se.vruntime value --- i.e., tasks would execute simultaneously and no task
  28. would ever get "out of balance" from the "ideal" share of CPU time. ]
  29. CFS's task picking logic is based on this p->se.vruntime value and it is thus
  30. very simple: it always tries to run the task with the smallest p->se.vruntime
  31. value (i.e., the task which executed least so far). CFS always tries to split
  32. up CPU time between runnable tasks as close to "ideal multitasking hardware" as
  33. possible.
  34. Most of the rest of CFS's design just falls out of this really simple concept,
  35. with a few add-on embellishments like nice levels, multiprocessing and various
  36. algorithm variants to recognize sleepers.
  37. 3. THE RBTREE
  38. ==============
  39. CFS's design is quite radical: it does not use the old data structures for the
  40. runqueues, but it uses a time-ordered rbtree to build a "timeline" of future
  41. task execution, and thus has no "array switch" artifacts (by which both the
  42. previous vanilla scheduler and RSDL/SD are affected).
  43. CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic
  44. increasing value tracking the smallest vruntime among all tasks in the
  45. runqueue. The total amount of work done by the system is tracked using
  46. min_vruntime; that value is used to place newly activated entities on the left
  47. side of the tree as much as possible.
  48. The total number of running tasks in the runqueue is accounted through the
  49. rq->cfs.load value, which is the sum of the weights of the tasks queued on the
  50. runqueue.
  51. CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the
  52. p->se.vruntime key. CFS picks the "leftmost" task from this tree and sticks to it.
  53. As the system progresses forwards, the executed tasks are put into the tree
  54. more and more to the right --- slowly but surely giving a chance for every task
  55. to become the "leftmost task" and thus get on the CPU within a deterministic
  56. amount of time.
  57. Summing up, CFS works like this: it runs a task a bit, and when the task
  58. schedules (or a scheduler tick happens) the task's CPU usage is "accounted
  59. for": the (small) time it just spent using the physical CPU is added to
  60. p->se.vruntime. Once p->se.vruntime gets high enough so that another task
  61. becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a
  62. small amount of "granularity" distance relative to the leftmost task so that we
  63. do not over-schedule tasks and trash the cache), then the new leftmost task is
  64. picked and the current task is preempted.
  65. 4. SOME FEATURES OF CFS
  66. ========================
  67. CFS uses nanosecond granularity accounting and does not rely on any jiffies or
  68. other HZ detail. Thus the CFS scheduler has no notion of "timeslices" in the
  69. way the previous scheduler had, and has no heuristics whatsoever. There is
  70. only one central tunable (you have to switch on CONFIG_SCHED_DEBUG):
  71. /proc/sys/kernel/sched_min_granularity_ns
  72. which can be used to tune the scheduler from "desktop" (i.e., low latencies) to
  73. "server" (i.e., good batching) workloads. It defaults to a setting suitable
  74. for desktop workloads. SCHED_BATCH is handled by the CFS scheduler module too.
  75. Due to its design, the CFS scheduler is not prone to any of the "attacks" that
  76. exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c,
  77. chew.c, ring-test.c, massive_intr.c all work fine and do not impact
  78. interactivity and produce the expected behavior.
  79. The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH
  80. than the previous vanilla scheduler: both types of workloads are isolated much
  81. more aggressively.
  82. SMP load-balancing has been reworked/sanitized: the runqueue-walking
  83. assumptions are gone from the load-balancing code now, and iterators of the
  84. scheduling modules are used. The balancing code got quite a bit simpler as a
  85. result.
  86. 5. Scheduling policies
  87. ======================
  88. CFS implements three scheduling policies:
  89. - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling
  90. policy that is used for regular tasks.
  91. - SCHED_BATCH: Does not preempt nearly as often as regular tasks
  92. would, thereby allowing tasks to run longer and make better use of
  93. caches but at the cost of interactivity. This is well suited for
  94. batch jobs.
  95. - SCHED_IDLE: This is even weaker than nice 19, but its not a true
  96. idle timer scheduler in order to avoid to get into priority
  97. inversion problems which would deadlock the machine.
  98. SCHED_FIFO/_RR are implemented in sched/rt.c and are as specified by
  99. POSIX.
  100. The command chrt from util-linux-ng 2.13.1.1 can set all of these except
  101. SCHED_IDLE.
  102. 6. SCHEDULING CLASSES
  103. ======================
  104. The new CFS scheduler has been designed in such a way to introduce "Scheduling
  105. Classes," an extensible hierarchy of scheduler modules. These modules
  106. encapsulate scheduling policy details and are handled by the scheduler core
  107. without the core code assuming too much about them.
  108. sched/fair.c implements the CFS scheduler described above.
  109. sched/rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than
  110. the previous vanilla scheduler did. It uses 100 runqueues (for all 100 RT
  111. priority levels, instead of 140 in the previous scheduler) and it needs no
  112. expired array.
  113. Scheduling classes are implemented through the sched_class structure, which
  114. contains hooks to functions that must be called whenever an interesting event
  115. occurs.
  116. This is the (partial) list of the hooks:
  117. - enqueue_task(...)
  118. Called when a task enters a runnable state.
  119. It puts the scheduling entity (task) into the red-black tree and
  120. increments the nr_running variable.
  121. - dequeue_task(...)
  122. When a task is no longer runnable, this function is called to keep the
  123. corresponding scheduling entity out of the red-black tree. It decrements
  124. the nr_running variable.
  125. - yield_task(...)
  126. This function is basically just a dequeue followed by an enqueue, unless the
  127. compat_yield sysctl is turned on; in that case, it places the scheduling
  128. entity at the right-most end of the red-black tree.
  129. - check_preempt_curr(...)
  130. This function checks if a task that entered the runnable state should
  131. preempt the currently running task.
  132. - pick_next_task(...)
  133. This function chooses the most appropriate task eligible to run next.
  134. - set_curr_task(...)
  135. This function is called when a task changes its scheduling class or changes
  136. its task group.
  137. - task_tick(...)
  138. This function is mostly called from time tick functions; it might lead to
  139. process switch. This drives the running preemption.
  140. 7. GROUP SCHEDULER EXTENSIONS TO CFS
  141. =====================================
  142. Normally, the scheduler operates on individual tasks and strives to provide
  143. fair CPU time to each task. Sometimes, it may be desirable to group tasks and
  144. provide fair CPU time to each such task group. For example, it may be
  145. desirable to first provide fair CPU time to each user on the system and then to
  146. each task belonging to a user.
  147. CONFIG_CGROUP_SCHED strives to achieve exactly that. It lets tasks to be
  148. grouped and divides CPU time fairly among such groups.
  149. CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and
  150. SCHED_RR) tasks.
  151. CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and
  152. SCHED_BATCH) tasks.
  153. These options need CONFIG_CGROUPS to be defined, and let the administrator
  154. create arbitrary groups of tasks, using the "cgroup" pseudo filesystem. See
  155. Documentation/admin-guide/cgroup-v1/cgroups.rst for more information about this filesystem.
  156. When CONFIG_FAIR_GROUP_SCHED is defined, a "cpu.shares" file is created for each
  157. group created using the pseudo filesystem. See example steps below to create
  158. task groups and modify their CPU share using the "cgroups" pseudo filesystem::
  159. # mount -t tmpfs cgroup_root /sys/fs/cgroup
  160. # mkdir /sys/fs/cgroup/cpu
  161. # mount -t cgroup -ocpu none /sys/fs/cgroup/cpu
  162. # cd /sys/fs/cgroup/cpu
  163. # mkdir multimedia # create "multimedia" group of tasks
  164. # mkdir browser # create "browser" group of tasks
  165. # #Configure the multimedia group to receive twice the CPU bandwidth
  166. # #that of browser group
  167. # echo 2048 > multimedia/cpu.shares
  168. # echo 1024 > browser/cpu.shares
  169. # firefox & # Launch firefox and move it to "browser" group
  170. # echo <firefox_pid> > browser/tasks
  171. # #Launch gmplayer (or your favourite movie player)
  172. # echo <movie_player_pid> > multimedia/tasks