From: Al Boldi <a1426z@gawab.com>
To: Con Kolivas <kernel@kolivas.org>
Cc: ck list <ck@vds.kolivas.org>,
linux-kernel@vger.kernel.org, Mike Galbraith <efault@gmx.de>
Subject: Re: [patch][rfc] quell interactive feeding frenzy
Date: Wed, 12 Apr 2006 13:39:40 +0300 [thread overview]
Message-ID: <200604121339.40563.a1426z@gawab.com> (raw)
In-Reply-To: <200604121936.16584.kernel@kolivas.org>
Con Kolivas wrote:
> On Wednesday 12 April 2006 18:17, Al Boldi wrote:
> > Con Kolivas wrote:
> > > Single heavily cpu bound computationally intensive tasks (think
> > > rendering etc).
> >
> > Why do you need a switch for that?
>
> Because avoiding doing need_resched and reassessing priority at less
> regular intervals means less overhead, and there is always something else
> running on a pc. At low loads the longer timeslices and delayed preemption
> contribute considerably to cache warmth and throughput. Comparing
> staircase's sched_compute mode on kernbench at "optimal loads" (make -j4 x
> num_cpus) showed the best throughput of all the schedulers tested.
Great!
> > > Sorry I don't understand what you mean. Why do you say it's not fair
> > > (got a testcase?). What do you mean by "definitely not smooth". What
> > > is smoothness and on what workloads is it not smooth? Also by ia you
> > > mean what?
> >
> > ia=interactivity i.e: responsiveness under high load.
> > smooth=not jumpy i.e: run '# gears & morph3d & reflect &' w/o stutter
>
> Installed and tested here just now. They run smoothly concurrently here.
> Are you testing on staircase15?
staircase14.2-test3. Are you testing w/ DRM? If not then all mesa requests
will be queued into X, and then runs as one task (check top d.1)
> > fair=non hogging i.e: spreading cpu-load across tasks evenly (top d.1)
>
> Only unblocked processes/threads where one depends on the other don't get
> equal share, which is as broken a testcase as relying on sched_yield. I
> have not seen a testcase demonstrating unfairness on current staircase.
> top shows me fair cpu usage.
Try ping -A (10x). top d.1 should show skewed times. If you have a fast
machine, you may have to increase the load.
> > > Again I don't understand. Just how heavy a load is heavy? Your
> > > testcases are already in what I would call stratospheric range. I
> > > don't personally think a cpu scheduler should be optimised for load
> > > infinity. And how are you defining efficient? You say it doesn't
> > > "look" efficient? What "looks" inefficient about it?
> >
> > The idea here is to expose inefficiencies by driving the system into
> > saturation, and although staircase is more efficient than the default
> > 2.6 scheduler, it is obviously less efficient than spa.
>
> Where do you stop calling something saturation and start calling it
> absurd? By your reckoning staircase is stable to loads of 300 on one cpu.
> spa being stable to higher loads is hardly comparable given the
> interactivity disparity between it and staircase. A compromise is one that
> does both very well; not one perfectly and the other poorly.
>
> > > You want tunables? The only tunable in staircase is rr_interval which
> > > (in -ck) has an on/off for big/small (sched_compute) since most other
> > > numbers in between (in my experience) are pretty meaningless. I could
> > > export rr_interval directly instead... I've not seen a good argument
> > > for doing that. Got one?
> >
> > Smoothness control, maybe?
>
> Have to think about that one. I'm not seeing a smoothness issue.
>
> > > However there are no other tunables at all (just look at
> > > the code). All tasks of any nice level have available the whole
> > > priority range from 100-139 which appears as PRIO 0-39 on top.
> > > Limiting that (again) changes the semantics.
> >
> > Yes, limiting this could change the semantics for the sake of fairness,
> > it's up to you.
>
> There is no problem with fairness that I am aware of.
Let's see after you retry the tests.
Thanks!
--
Al
next prev parent reply other threads:[~2006-04-12 10:41 UTC|newest]
Thread overview: 54+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <200604112100.28725.kernel@kolivas.org>
2006-04-11 17:03 ` Fwd: Re: [patch][rfc] quell interactive feeding frenzy Al Boldi
2006-04-11 22:56 ` Con Kolivas
2006-04-12 5:41 ` Al Boldi
2006-04-12 6:22 ` Con Kolivas
2006-04-12 8:17 ` Al Boldi
2006-04-12 9:36 ` Con Kolivas
2006-04-12 10:39 ` Al Boldi [this message]
2006-04-12 11:27 ` Con Kolivas
2006-04-12 15:25 ` Al Boldi
2006-04-13 11:51 ` Con Kolivas
2006-04-14 3:16 ` Al Boldi
2006-04-15 7:05 ` Con Kolivas
2006-04-15 18:23 ` [ck] " Michael Gerdau
2006-04-15 20:45 ` Al Boldi
2006-04-15 23:22 ` Con Kolivas
2006-04-16 18:44 ` [ck] " Andreas Mohr
2006-04-17 0:08 ` Con Kolivas
2006-04-19 8:37 ` Andreas Mohr
2006-04-19 8:59 ` jos poortvliet
2006-04-15 22:32 ` jos poortvliet
2006-04-15 23:06 ` Con Kolivas
2006-04-16 6:02 ` Con Kolivas
2006-04-16 8:31 ` Al Boldi
2006-04-16 8:58 ` Con Kolivas
2006-04-16 10:37 ` was " Con Kolivas
2006-04-16 19:03 ` Al Boldi
2006-04-16 23:26 ` Con Kolivas
2006-04-09 16:44 [patch][rfc] " Al Boldi
2006-04-09 18:33 ` Mike Galbraith
2006-04-10 14:43 ` Al Boldi
2006-04-11 10:57 ` Con Kolivas
-- strict thread matches above, loose matches on Subject: below --
2006-04-07 9:38 Mike Galbraith
2006-04-07 9:47 ` Andrew Morton
2006-04-07 9:52 ` Ingo Molnar
2006-04-07 10:57 ` Mike Galbraith
2006-04-07 11:00 ` Con Kolivas
2006-04-07 11:09 ` Mike Galbraith
2006-04-07 10:40 ` Mike Galbraith
2006-04-07 12:56 ` Con Kolivas
2006-04-07 13:37 ` Mike Galbraith
2006-04-07 13:56 ` Con Kolivas
2006-04-07 14:14 ` Mike Galbraith
2006-04-07 15:16 ` Mike Galbraith
2006-04-09 11:14 ` bert hubert
2006-04-09 11:39 ` Mike Galbraith
2006-04-09 12:14 ` bert hubert
2006-04-09 18:07 ` Mike Galbraith
2006-04-10 9:12 ` bert hubert
2006-04-10 10:00 ` Mike Galbraith
2006-04-10 14:56 ` Mike Galbraith
2006-04-13 7:41 ` Mike Galbraith
2006-04-13 10:16 ` Con Kolivas
2006-04-13 11:05 ` Mike Galbraith
2006-04-09 18:24 ` Mike Galbraith
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=200604121339.40563.a1426z@gawab.com \
--to=a1426z@gawab.com \
--cc=ck@vds.kolivas.org \
--cc=efault@gmx.de \
--cc=kernel@kolivas.org \
--cc=linux-kernel@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).