* William Lee Irwin III <wli@holomorphy.com> wrote:

quoted text
> Where it gets complex is when the behavior patterns vary, e.g. they're 
> not entirely CPU-bound and their desired in-isolation CPU utilization 
> varies, or when nice levels vary, or both vary. [...]

yes. I tested things like 'massive_intr.c' (attached, written by Satoru 
Takeuchi) which starts N tasks which each work for 8msec then sleep 
1msec:

from its output, the second column is the CPU time each thread got, the 
more even, the fairer the scheduling. On vanilla i get:

 mercury:~> ./massive_intr 10 10
 024873  00000150
 024874  00000123
 024870  00000069
 024868  00000068
 024866  00000051
 024875  00000206
 024872  00000093
 024869  00000138
 024867  00000078
 024871  00000223

on CFS i get:

 neptune:~> ./massive_intr 10 10
 002266  00000112
 002260  00000113
 002261  00000112
 002267  00000112
 002269  00000112
 002265  00000112
 002262  00000113
 002268  00000113
 002264  00000112
 002263  00000113

so it is quite a bit more even ;)

another related test-utility is one i wrote:

  http://people.redhat.com/mingo/scheduler-patches/ring-test.c

this is a ring of 100 tasks each doing work for 100 msecs and then 
sleeping for 1 msec. I usually test this by also running a CPU hog in 
parallel to it, and checking whether it gets ~50.0% of CPU time under 
CFS. (it does)

	Ingo