More fuel on the BFS vs. CFS fire

#1

Post by DarkShikari » Tue, Sep 15, 2009, 4:55 AM UTC

More fuel on the BFS vs. CFS fire
thread.gmane.org

Re: More fuel on the BFS vs. CFS fire

#2

Post by ars » Tue, Sep 15, 2009, 9:54 AM UTC

Background:

This is about the choice of CPU scheduler in the linux kernel.

CFS: Completely Fair Scheduling, current default.

BFS: Brain Fuck Scheduler, written by the infamous Con Kolivas, see http://news.ycombinator.com/item?id=799147

Re: More fuel on the BFS vs. CFS fire

#3

Post by viraptor » Tue, Sep 15, 2009, 10:20 AM UTC

"fuel"? It's the first (that I've seen) well described, reproducible test that includes a howto for retesting in this debate... Whether someone improves BFS (cpu jumping) or CFS (general performance) because of this post, we win. Great!

Edit: And that's what happened (more or less) - in the first response:

2.6.31-bfs221-smp - 408.38 fps

2.6.32-tip-smp NO_NEXT_BUDDY NO_LB_BIAS - 418.07 fps

Re: More fuel on the BFS vs. CFS fire

#4

Post by ars » Tue, Sep 15, 2009, 10:43 AM UTC

viraptor wrote: post #3

"fuel"? It's the first (that I've seen) well described, reproducible test that includes a howto for retesting in this debate... Whether someone improves BFS (cpu jumping) or CFS (general performance) because of this post, we win. Great! Edit: And that's what happened (more or less) - in the first response: 2.6.31-bfs221-smp - 408.38 fps 2.6.32-tip-smp NO_NEXT_BUDDY NO_LB_BIAS - 418.07 fps

It's not the first, there is also: http://lwn.net/Articles/351058/

But, I believe, that test suffers from one fatal flaw: Hyperthreading was enabled, and as far as I know BFS does not support hyperthreading. (Support means: know that that the two CPUs are really the same, rather than treat them as if they were two different physical CPUs.)

Re: More fuel on the BFS vs. CFS fire

#5

Post by viraptor » Tue, Sep 15, 2009, 10:54 AM UTC

ars wrote: post #4

viraptor wrote: post #3

"fuel"? It's the first (that I've seen) well described, reproducible test that includes a howto for retesting in this debate... Whether someone improves BFS (cpu jumping) or CFS (general performance) because of this post, we win. Great! Edit: And that's what happened (more or less) - in the first response: 2.6.31-bfs221-smp - 408.38 fps 2.6.32-tip-smp NO_NEXT_BUDDY NO_LB_BIAS - 418.07 fps

It's not the first, there is also: http://lwn.net/Articles/351058/ But, I believe, that test suffers from one fatal flaw: Hyperthreading was enabled, and as far as I know BFS does not support hyperthreading. (Support means: know that that the two CPUs are really the same, rather than treat them as if they were two different physical CPUs.)

Ah yeah - sorry - I'm just deliberately ignoring the "official" benchmarks from Ingo and Con... At least until one of them publishes a benchmark that shows their own solution loses in some scenario ;)

Re: More fuel on the BFS vs. CFS fire

#6

Post by asyazwan » Wed, Sep 16, 2009, 9:15 AM UTC

ars wrote: post #2

Background: This is about the choice of CPU scheduler in the linux kernel. CFS: Completely Fair Scheduling, current default. BFS: Brain Fuck Scheduler, written by the infamous Con Kolivas, see http://news.ycombinator.com/item?id=799147

Is that name serious..?

EDIT: I know a language named similar but that's more like a for-fun project, not a serious one to the extent of replacing existing solutions