Re: Linux scheduler, overscheduling performance, threads

[Horst von Brand]

Brian Hurt <[e-mail]> said:
> This discussion is rapidly degenerating into a "does too!" "does not!"
> match.  So let me turn the question around.  Assume, for a moment, that a
> patch for Linux existed which did two things:
> 1) Ran signifigantly faster for large run queues- for example it switched
> from the current O(n) algorithm to an O(log n) algorithm.
> 2) Ran slower than the current scheduling algorithm for short run
> queues.
> And the patch had no other effects on the kernel at large and was
> otherwise well written, etc.

> Am I right in assuming that wether the patch would be accepted would
> depend upon how much slower it made the common case?   Obviously, if it
> made the common case no slower, no one would mind putting it in the
> kernel.  On the other hand, if it made the common case a million times
> slower, there is no way it'd ever get into the kernel, and rightfully so.
> Where (roughly) inbetween is the breakpoint?

When it doesn't make runqueues of length 1 (typical, _optimal_ case) slower
at all. OTOH, this happens once a timeslice, let's say 5 to 100 times a
second, so it is simply not performance critical
-- 
Horst von Brand                             [e-mail]
Casilla 9G, Viņa del Mar, Chile                               [phone]