LWN: Comments on "The Rotating Staircase Deadline Scheduler"

The Rotating Staircase Deadline Scheduler

muwlgr — Sat, 17 Mar 2007 08:00:42 +0000

On my system, KDE's 3.5.5 Konqueror uses CPU time in a very strange way. It sleeps for about 60% of time, then does something CPU-intensive for 40%. Its interchanging sleeps/runs are quite short, so its priority is not lowered by the scheduler, and its CPU consumption has its visible impact on Mozilla Seamonkey 1.1.1 running in the nearby X session on the same CPU. Manual renicing of Konqueror to level 19 helps, but should not it be automatic ?

I wonder if this new scheduler can detect such "isochronous" CPU consumers and deal with them accordingly ?

The Rotating Staircase Deadline Scheduler

guest — Thu, 15 Mar 2007 17:27:54 +0000

I thought Con had said this patch is better for server loads, not desktops. How does this compare with his Staircase scheduler for interactivity?

The Rotating Staircase Deadline Scheduler

jd — Thu, 15 Mar 2007 16:58:21 +0000

Dimensionally Transcendent Lorentzian Transformational Scheduler

Huh. A scheduler that's larger on the inside than the outside. You do realize that you'll have to return the book you copied it from to the Panopticon Library on Gallifrey?

At one point, HP produced a scheduler plugin system. If anyone has a copy of that still, there may be ideas in it worth plundering, err re-using, as no algorithm is going to be good for all cases. What you want is a hypervisor of some sort to analyze the generalized characteristics and swap scheduler if the one currently running is unsuitable for the problem at hand.

The Rotating Staircase Deadline Scheduler

aigarius — Tue, 13 Mar 2007 17:56:50 +0000

Finally! A scheduler that non-kerlen programmers can understand with ease. Pure genius.

The Rotating Staircase Deadline Scheduler

njs — Thu, 08 Mar 2007 23:49:05 +0000

>Dimensionally Transcendent Lorentzian Transformational Scheduler

Is that the one that lets processes perform an infinite loop so fast that they travel back in time and become their own process group leader?

Self-renicing emerge

vmole — Thu, 08 Mar 2007 19:46:58 +0000

Apt and dpkg are not, in general, CPU bound, but I/O bound. Nicing I/O bound processes generally causes more problems than it solves, because you end up blowing the file cache.

The Rotating Staircase Deadline Scheduler

sbergman27 — Thu, 08 Mar 2007 19:10:14 +0000

"""CPU scheduling seems to be one of those eternally unfinished jobs."""

Yes, indeed. But every time I see an article like this, I can't help but think back to that time, some years ago, when interest in the Linux process scheduler really fired up. It may have been during 2.1.x, but I can't remember for sure.

I do remember Linus stating on LKML that he didn't think that process scheduling was very interesting. In his opinion, process scheduling was the sort of thing that you worked on, got right, and then left for a project that was actually challenging.

I wonder how he might have responded to a proposal called the "Rotating Staircase DeadLine Scheduler" back then?

At any rate, RSDL is doomed once I finish up and present my Dimensionally Transcendent Lorentzian Transformational Scheduler. It's almost finished. But seems to be stuck in a loop at the moment... ;-)

Self-renicing emerge

farnz — Thu, 08 Mar 2007 14:35:46 +0000

Although it's not currently on by default, emerge will already renice itself if you set the PORTAGE_NICENESS variable in /etc/make.conf. I do this so that an emerge doesn't kill system performance.

I've not prodded Debian for a while, but I'd be surprised if there wasn't a similar setting for apt/dpkg.

The Rotating Staircase Deadline Scheduler

dion — Thu, 08 Mar 2007 12:54:22 +0000

I'd like to second that idea about selfnicing, but perhaps it would be better to simply let crond do that?

The Rotating Staircase Deadline Scheduler

jospoortvliet — Thu, 08 Mar 2007 10:47:09 +0000

Lovely article ;-)

It's good to see this getting some attention, as RSDL seems to be a good piece of work. Starvation has been a problem in the kernel, giving short stalls now and then. A completely fair yet interactive scheduler like this one would do away with that, at the expense of ppl having to use nice probably a bit more on heavy processes like compiling.

Maybe apps like dpkg or emerge should start nicing themselves by default...

Still, RSDL gives a perfectly responsive desktop even when a make -j4 is running aside with aMule, mail etcetera, so its doing better than mainline on my system already.

And a better throughput is really an unexpected but nice benefit. Maybe worth mentioning RSDL does a bit better than mainline on the MySQL scaling issue:
http://jeffr-tech.livejournal.com/5705.html
http://bhhdoa.org.au/pipermail/ck/2007-March/006790.html
http://bhhdoa.org.au/pipermail/ck/2007-March/006794.html