aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation')
-rw-r--r--Documentation/sched-design-CFS.txt2
-rw-r--r--Documentation/sched-nice-design.txt108
2 files changed, 109 insertions, 1 deletions
diff --git a/Documentation/sched-design-CFS.txt b/Documentation/sched-design-CFS.txt
index 16feebb7bdc0..84901e7c0508 100644
--- a/Documentation/sched-design-CFS.txt
+++ b/Documentation/sched-design-CFS.txt
@@ -83,7 +83,7 @@ Some implementation details:
83 CFS uses nanosecond granularity accounting and does not rely on any 83 CFS uses nanosecond granularity accounting and does not rely on any
84 jiffies or other HZ detail. Thus the CFS scheduler has no notion of 84 jiffies or other HZ detail. Thus the CFS scheduler has no notion of
85 'timeslices' and has no heuristics whatsoever. There is only one 85 'timeslices' and has no heuristics whatsoever. There is only one
86 central tunable: 86 central tunable (you have to switch on CONFIG_SCHED_DEBUG):
87 87
88 /proc/sys/kernel/sched_granularity_ns 88 /proc/sys/kernel/sched_granularity_ns
89 89
diff --git a/Documentation/sched-nice-design.txt b/Documentation/sched-nice-design.txt
new file mode 100644
index 000000000000..e2bae5a577e3
--- /dev/null
+++ b/Documentation/sched-nice-design.txt
@@ -0,0 +1,108 @@
1This document explains the thinking about the revamped and streamlined
2nice-levels implementation in the new Linux scheduler.
3
4Nice levels were always pretty weak under Linux and people continuously
5pestered us to make nice +19 tasks use up much less CPU time.
6
7Unfortunately that was not that easy to implement under the old
8scheduler, (otherwise we'd have done it long ago) because nice level
9support was historically coupled to timeslice length, and timeslice
10units were driven by the HZ tick, so the smallest timeslice was 1/HZ.
11
12In the O(1) scheduler (in 2003) we changed negative nice levels to be
13much stronger than they were before in 2.4 (and people were happy about
14that change), and we also intentionally calibrated the linear timeslice
15rule so that nice +19 level would be _exactly_ 1 jiffy. To better
16understand it, the timeslice graph went like this (cheesy ASCII art
17alert!):
18
19
20 A
21 \ | [timeslice length]
22 \ |
23 \ |
24 \ |
25 \ |
26 \|___100msecs
27 |^ . _
28 | ^ . _
29 | ^ . _
30 -*----------------------------------*-----> [nice level]
31 -20 | +19
32 |
33 |
34
35So that if someone wanted to really renice tasks, +19 would give a much
36bigger hit than the normal linear rule would do. (The solution of
37changing the ABI to extend priorities was discarded early on.)
38
39This approach worked to some degree for some time, but later on with
40HZ=1000 it caused 1 jiffy to be 1 msec, which meant 0.1% CPU usage which
41we felt to be a bit excessive. Excessive _not_ because it's too small of
42a CPU utilization, but because it causes too frequent (once per
43millisec) rescheduling. (and would thus trash the cache, etc. Remember,
44this was long ago when hardware was weaker and caches were smaller, and
45people were running number crunching apps at nice +19.)
46
47So for HZ=1000 we changed nice +19 to 5msecs, because that felt like the
48right minimal granularity - and this translates to 5% CPU utilization.
49But the fundamental HZ-sensitive property for nice+19 still remained,
50and we never got a single complaint about nice +19 being too _weak_ in
51terms of CPU utilization, we only got complaints about it (still) being
52too _strong_ :-)
53
54To sum it up: we always wanted to make nice levels more consistent, but
55within the constraints of HZ and jiffies and their nasty design level
56coupling to timeslices and granularity it was not really viable.
57
58The second (less frequent but still periodically occuring) complaint
59about Linux's nice level support was its assymetry around the origo
60(which you can see demonstrated in the picture above), or more
61accurately: the fact that nice level behavior depended on the _absolute_
62nice level as well, while the nice API itself is fundamentally
63"relative":
64
65 int nice(int inc);
66
67 asmlinkage long sys_nice(int increment)
68
69(the first one is the glibc API, the second one is the syscall API.)
70Note that the 'inc' is relative to the current nice level. Tools like
71bash's "nice" command mirror this relative API.
72
73With the old scheduler, if you for example started a niced task with +1
74and another task with +2, the CPU split between the two tasks would
75depend on the nice level of the parent shell - if it was at nice -10 the
76CPU split was different than if it was at +5 or +10.
77
78A third complaint against Linux's nice level support was that negative
79nice levels were not 'punchy enough', so lots of people had to resort to
80run audio (and other multimedia) apps under RT priorities such as
81SCHED_FIFO. But this caused other problems: SCHED_FIFO is not starvation
82proof, and a buggy SCHED_FIFO app can also lock up the system for good.
83
84The new scheduler in v2.6.23 addresses all three types of complaints:
85
86To address the first complaint (of nice levels being not "punchy"
87enough), the scheduler was decoupled from 'time slice' and HZ concepts
88(and granularity was made a separate concept from nice levels) and thus
89it was possible to implement better and more consistent nice +19
90support: with the new scheduler nice +19 tasks get a HZ-independent
911.5%, instead of the variable 3%-5%-9% range they got in the old
92scheduler.
93
94To address the second complaint (of nice levels not being consistent),
95the new scheduler makes nice(1) have the same CPU utilization effect on
96tasks, regardless of their absolute nice levels. So on the new
97scheduler, running a nice +10 and a nice 11 task has the same CPU
98utilization "split" between them as running a nice -5 and a nice -4
99task. (one will get 55% of the CPU, the other 45%.) That is why nice
100levels were changed to be "multiplicative" (or exponential) - that way
101it does not matter which nice level you start out from, the 'relative
102result' will always be the same.
103
104The third complaint (of negative nice levels not being "punchy" enough
105and forcing audio apps to run under the more dangerous SCHED_FIFO
106scheduling policy) is addressed by the new scheduler almost
107automatically: stronger negative nice levels are an automatic
108side-effect of the recalibrated dynamic range of nice levels.