Messages in this thread | | | Date | Sat, 11 Sep 2010 16:52:14 -0400 | From | Mathieu Desnoyers <> | Subject | Re: [RFC patch 1/2] sched: dynamically adapt granularity with nr_running |
| |
* Peter Zijlstra (peterz@infradead.org) wrote: > On Sat, 2010-09-11 at 12:21 -0700, Linus Torvalds wrote: > > On Sat, Sep 11, 2010 at 11:57 AM, Peter Zijlstra <peterz@infradead.org> wrote: > > > > > > Not at all charmed, this look like random changes without conceptual > > > integrity. > > > > I wish people actually looked at the _numbers_ and reacted to them, > > rather than argue theory. > > > > Guys, we have cases of bad latency under load. That's a pretty > > undeniable fact. Arguing against a patch because of some theoretical > > issue without at all even acknowledging the latency improvements is, I > > think, really bad form. > > > > So please. Acknowledge the latency issue. And come up with better > > patches, rather than just shoot down alternatives. Because if the > > answer is just NAK with no alternative, then that answer is worthless. > > No? > > >From what I can make up: > > LAT=`cat /proc/sys/kernel/sched_latency_ns`; > echo $((LAT/8)) > /proc/sys/kernel/sched_min_granularity_ns > > will give you pretty much the same result as Mathieu's patch.
Not quite. Doing what you propose here would change the scheduling granularity (thus decrease throughput because it would schedule more often) when there are few tasks running. My approach does not: it only shrinks granularity when the number of running tasks increases over 3.
Thanks,
Mathieu
> > But if you want us to change the scheduler to be more latency sensitive > and trade in throughput for other benchmarks, we can do that.
-- Mathieu Desnoyers Operating System Efficiency R&D Consultant EfficiOS Inc. http://www.efficios.com
| |