On Sun, 2011-01-30 at 16:04 +0200, Avi Kivity wrote:
> On 01/28/2011 09:52 PM, Glauber Costa wrote:
> > This patch accounts steal time time in kernel/sched.
> > I kept it from last proposal, because I still see advantages
> > in it: Doing it here will give us easier access from scheduler
> > variables such as the cpu rq. The next patch shows an example of
> > usage for it.
> >
> > Since functions like account_idle_time() can be called from
> > multiple places, not only account_process_tick(), steal time
> > grabbing is repeated in each account function separatedely.
> >
>
> I accept that steal time is worthwhile, but do you have some way to
> demonstrate that the implementation actually works and is beneficial?
>
> Perhaps run two cpu-bound compute processes on one vcpu, overcommit that
> vcpu, and see what happens to the processing rate with and without steal
> time accounting. I'd expect a fairer response with steal time accounting.
Avi,
There are two things here:
One of them, which is solely the accounting of steal time, (patches 1 to
4) has absolutely nothing to do with what you said. Its sole purpose is
to provide the user with information about "why is my process slow if I
am using 100 % of my cpu?")
The last patch is the only one that actually tries to rebalance cpus
according to steal time information. For that, I do have some
experiments I did here to see if it is working, will try to provide more
precise data in the next submission.
On 02/01/2011 05:57 PM, Glauber Costa wrote:
> On Sun, 2011-01-30 at 16:04 +0200, Avi Kivity wrote:
> > On 01/28/2011 09:52 PM, Glauber Costa wrote:
> > > This patch accounts steal time time in kernel/sched.
> > > I kept it from last proposal, because I still see advantages
> > > in it: Doing it here will give us easier access from scheduler
> > > variables such as the cpu rq. The next patch shows an example of
> > > usage for it.
> > >
> > > Since functions like account_idle_time() can be called from
> > > multiple places, not only account_process_tick(), steal time
> > > grabbing is repeated in each account function separatedely.
> > >
> >
> > I accept that steal time is worthwhile, but do you have some way to
> > demonstrate that the implementation actually works and is beneficial?
> >
> > Perhaps run two cpu-bound compute processes on one vcpu, overcommit that
> > vcpu, and see what happens to the processing rate with and without steal
> > time accounting. I'd expect a fairer response with steal time accounting.
>
> Avi,
>
> There are two things here:
> One of them, which is solely the accounting of steal time, (patches 1 to
> 4) has absolutely nothing to do with what you said. Its sole purpose is
> to provide the user with information about "why is my process slow if I
> am using 100 % of my cpu?")
Right. Like irq and softirq time, we need to report this to the user,
as it's potentially much higher.
> The last patch is the only one that actually tries to rebalance cpus
> according to steal time information. For that, I do have some
> experiments I did here to see if it is working, will try to provide more
> precise data in the next submission.
>
Thanks.
--
error compiling committee.c: too many arguments to function
On 02/02/2011 12:11 PM, Avi Kivity wrote:
> On 02/01/2011 05:57 PM, Glauber Costa wrote:
>> On Sun, 2011-01-30 at 16:04 +0200, Avi Kivity wrote:
>> > On 01/28/2011 09:52 PM, Glauber Costa wrote:
>> > > This patch accounts steal time time in kernel/sched.
>> > > I kept it from last proposal, because I still see advantages
>> > > in it: Doing it here will give us easier access from scheduler
>> > > variables such as the cpu rq. The next patch shows an example of
>> > > usage for it.
>> > >
>> > > Since functions like account_idle_time() can be called from
>> > > multiple places, not only account_process_tick(), steal time
>> > > grabbing is repeated in each account function separatedely.
>> > >
>> >
>> > I accept that steal time is worthwhile, but do you have some way to
>> > demonstrate that the implementation actually works and is beneficial?
>> >
>> > Perhaps run two cpu-bound compute processes on one vcpu,
>> overcommit that
>> > vcpu, and see what happens to the processing rate with and without
>> steal
>> > time accounting. I'd expect a fairer response with steal time
>> accounting.
>>
>> Avi,
>>
>> There are two things here:
>> One of them, which is solely the accounting of steal time, (patches 1 to
>> 4) has absolutely nothing to do with what you said. Its sole purpose is
>> to provide the user with information about "why is my process slow if I
>> am using 100 % of my cpu?")
>
> Right. Like irq and softirq time, we need to report this to the user,
> as it's potentially much higher.
Of course, it's not enough to just account for this time, you also have
to expose it somewhere, and update tools like top(1) to display it.
--
error compiling committee.c: too many arguments to function
On Wed, 2011-02-02 at 12:51 +0200, Avi Kivity wrote:
> On 02/02/2011 12:11 PM, Avi Kivity wrote:
> > On 02/01/2011 05:57 PM, Glauber Costa wrote:
> >> On Sun, 2011-01-30 at 16:04 +0200, Avi Kivity wrote:
> >> > On 01/28/2011 09:52 PM, Glauber Costa wrote:
> >> > > This patch accounts steal time time in kernel/sched.
> >> > > I kept it from last proposal, because I still see advantages
> >> > > in it: Doing it here will give us easier access from scheduler
> >> > > variables such as the cpu rq. The next patch shows an example of
> >> > > usage for it.
> >> > >
> >> > > Since functions like account_idle_time() can be called from
> >> > > multiple places, not only account_process_tick(), steal time
> >> > > grabbing is repeated in each account function separatedely.
> >> > >
> >> >
> >> > I accept that steal time is worthwhile, but do you have some way to
> >> > demonstrate that the implementation actually works and is beneficial?
> >> >
> >> > Perhaps run two cpu-bound compute processes on one vcpu,
> >> overcommit that
> >> > vcpu, and see what happens to the processing rate with and without
> >> steal
> >> > time accounting. I'd expect a fairer response with steal time
> >> accounting.
> >>
> >> Avi,
> >>
> >> There are two things here:
> >> One of them, which is solely the accounting of steal time, (patches 1 to
> >> 4) has absolutely nothing to do with what you said. Its sole purpose is
> >> to provide the user with information about "why is my process slow if I
> >> am using 100 % of my cpu?")
> >
> > Right. Like irq and softirq time, we need to report this to the user,
> > as it's potentially much higher.
>
> Of course, it's not enough to just account for this time, you also have
> to expose it somewhere, and update tools like top(1) to display it.
Yes, what I meant is that just the accounting will just expose it to the
tools, won't affect the scheduler in any sense.