| blob | 1605bf0cba8bd5da6b9f81502e7256094056d9bb |
1 Goals, Design and Implementation of the
2 new ultra-scalable O(1) scheduler
5 This is an edited version of an email Ingo Molnar sent to
6 lkml on 4 Jan 2002. It describes the goals, design, and
7 implementation of Ingo's new ultra-scalable O(1) scheduler.
8 Last Updated: 18 April 2002.
11 Goal
12 ====
14 The main goal of the new scheduler is to keep all the good things we know
15 and love about the current Linux scheduler:
17 - good interactive performance even during high load: if the user
18 types or clicks then the system must react instantly and must execute
19 the user tasks smoothly, even during considerable background load.
21 - good scheduling/wakeup performance with 1-2 runnable processes.
23 - fairness: no process should stay without any timeslice for any
24 unreasonable amount of time. No process should get an unjustly high
25 amount of CPU time.
27 - priorities: less important tasks can be started with lower priority,
28 more important tasks with higher priority.
30 - SMP efficiency: no CPU should stay idle if there is work to do.
32 - SMP affinity: processes which run on one CPU should stay affine to
33 that CPU. Processes should not bounce between CPUs too frequently.
35 - plus additional scheduler features: RT scheduling, CPU binding.
37 and the goal is also to add a few new things:
39 - fully O(1) scheduling. Are you tired of the recalculation loop
40 blowing the L1 cache away every now and then? Do you think the goodness
41 loop is taking a bit too long to finish if there are lots of runnable
42 processes? This new scheduler takes no prisoners: wakeup(), schedule(),
43 the timer interrupt are all O(1) algorithms. There is no recalculation
44 loop. There is no goodness loop either.
46 - 'perfect' SMP scalability. With the new scheduler there is no 'big'
47 runqueue_lock anymore - it's all per-CPU runqueues and locks - two
48 tasks on two separate CPUs can wake up, schedule and context-switch
49 completely in parallel, without any interlocking. All
50 scheduling-relevant data is structured for maximum scalability.
52 - better SMP affinity. The old scheduler has a particular weakness that
53 causes the random bouncing of tasks between CPUs if/when higher
54 priority/interactive tasks, this was observed and reported by many
55 people. The reason is that the timeslice recalculation loop first needs
56 every currently running task to consume its timeslice. But when this
57 happens on eg. an 8-way system, then this property starves an
58 increasing number of CPUs from executing any process. Once the last
59 task that has a timeslice left has finished using up that timeslice,
60 the recalculation loop is triggered and other CPUs can start executing
61 tasks again - after having idled around for a number of timer ticks.
62 The more CPUs, the worse this effect.
64 Furthermore, this same effect causes the bouncing effect as well:
65 whenever there is such a 'timeslice squeeze' of the global runqueue,
66 idle processors start executing tasks which are not affine to that CPU.
67 (because the affine tasks have finished off their timeslices already.)
69 The new scheduler solves this problem by distributing timeslices on a
70 per-CPU basis, without having any global synchronization or
71 recalculation.
73 - batch scheduling. A significant proportion of computing-intensive tasks
74 benefit from batch-scheduling, where timeslices are long and processes
75 are roundrobin scheduled. The new scheduler does such batch-scheduling
76 of the lowest priority tasks - so nice +19 jobs will get
77 'batch-scheduled' automatically. With this scheduler, nice +19 jobs are
78 in essence SCHED_IDLE, from an interactiveness point of view.
80 - handle extreme loads more smoothly, without breakdown and scheduling
81 storms.
83 - O(1) RT scheduling. For those RT folks who are paranoid about the
84 O(nr_running) property of the goodness loop and the recalculation loop.
86 - run fork()ed children before the parent. Andrea has pointed out the
87 advantages of this a few months ago, but patches for this feature
88 do not work with the old scheduler as well as they should,
89 because idle processes often steal the new child before the fork()ing
90 CPU gets to execute it.
93 Design
94 ======
96 The core of the new scheduler contains the following mechanisms:
98 - *two* priority-ordered 'priority arrays' per CPU. There is an 'active'
99 array and an 'expired' array. The active array contains all tasks that
100 are affine to this CPU and have timeslices left. The expired array
101 contains all tasks which have used up their timeslices - but this array
102 is kept sorted as well. The active and expired array is not accessed
103 directly, it's accessed through two pointers in the per-CPU runqueue
104 structure. If all active tasks are used up then we 'switch' the two
105 pointers and from now on the ready-to-go (former-) expired array is the
106 active array - and the empty active array serves as the new collector
107 for expired tasks.
109 - there is a 64-bit bitmap cache for array indices. Finding the highest
110 priority task is thus a matter of two x86 BSFL bit-search instructions.
112 the split-array solution enables us to have an arbitrary number of active
113 and expired tasks, and the recalculation of timeslices can be done
114 immediately when the timeslice expires. Because the arrays are always
115 access through the pointers in the runqueue, switching the two arrays can
116 be done very quickly.
118 this is a hybride priority-list approach coupled with roundrobin
119 scheduling and the array-switch method of distributing timeslices.
121 - there is a per-task 'load estimator'.
123 one of the toughest things to get right is good interactive feel during
124 heavy system load. While playing with various scheduler variants i found
125 that the best interactive feel is achieved not by 'boosting' interactive
126 tasks, but by 'punishing' tasks that want to use more CPU time than there
127 is available. This method is also much easier to do in an O(1) fashion.
129 to establish the actual 'load' the task contributes to the system, a
130 complex-looking but pretty accurate method is used: there is a 4-entry
131 'history' ringbuffer of the task's activities during the last 4 seconds.
132 This ringbuffer is operated without much overhead. The entries tell the
133 scheduler a pretty accurate load-history of the task: has it used up more
134 CPU time or less during the past N seconds. [the size '4' and the interval
135 of 4x 1 seconds was found by lots of experimentation - this part is
136 flexible and can be changed in both directions.]
138 the penalty a task gets for generating more load than the CPU can handle
139 is a priority decrease - there is a maximum amount to this penalty
140 relative to their static priority, so even fully CPU-bound tasks will
141 observe each other's priorities, and will share the CPU accordingly.
143 the SMP load-balancer can be extended/switched with additional parallel
144 computing and cache hierarchy concepts: NUMA scheduling, multi-core CPUs
145 can be supported easily by changing the load-balancer. Right now it's
146 tuned for my SMP systems.
148 i skipped the prev->mm == next->mm advantage - no workload i know of shows
149 any sensitivity to this. It can be added back by sacrificing O(1)
150 schedule() [the current and one-lower priority list can be searched for a
151 that->mm == current->mm condition], but costs a fair number of cycles
152 during a number of important workloads, so i wanted to avoid this as much
153 as possible.
155 - the SMP idle-task startup code was still racy and the new scheduler
156 triggered this. So i streamlined the idle-setup code a bit. We do not call
157 into schedule() before all processors have started up fully and all idle
158 threads are in place.
160 - the patch also cleans up a number of aspects of sched.c - moves code
161 into other areas of the kernel where it's appropriate, and simplifies
162 certain code paths and data constructs. As a result, the new scheduler's
163 code is smaller than the old one.
165 Ingo