On Thu, 2007-11-01 at 17:01 +0530, Srivatsa Vaddagiri wrote:
quoted text > On Wed, Oct 31, 2007 at 10:10:32PM +0100, Peter Zijlstra wrote:
> > Currently the ideal slice length does not take group scheduling into ac=
count.
quoted text > > Change it so that it properly takes all the runnable tasks on this cpu =
into
quoted text > > account and caluclate the weight according to the grouping hierarchy.
> >=20
> > Also fixes a bug in vslice which missed a factor NICE_0_LOAD.
> >=20
> > Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > CC: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
> > ---
> > kernel/sched_fair.c | 42 +++++++++++++++++++++++++++++++-----------
> > 1 file changed, 31 insertions(+), 11 deletions(-)
> >=20
> > Index: linux-2.6/kernel/sched_fair.c
> > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
quoted text > > --- linux-2.6.orig/kernel/sched_fair.c
> > +++ linux-2.6/kernel/sched_fair.c
> > @@ -331,10 +331,15 @@ static u64 __sched_period(unsigned long=20
> > */
> > static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
> > {
> > - u64 slice =3D __sched_period(cfs_rq->nr_running);
> > + unsigned long nr_running =3D rq_of(cfs_rq)->nr_running;
> > + u64 slice =3D __sched_period(nr_running);
> >=20
> > - slice *=3D se->load.weight;
> > - do_div(slice, cfs_rq->load.weight);
> > + for_each_sched_entity(se) {
> > + cfs_rq =3D cfs_rq_of(se);
> > +
> > + slice *=3D se->load.weight;
> > + do_div(slice, cfs_rq->load.weight);
> > + }
> >=20
> > return slice;
>=20
>=20
> Lets say we have two groups A and B on CPU0, of equal weight (1024).
>=20
> Further,
>=20
> A has 1 task (A0)
> B has 1000 tasks (B0 .. B999)=20
>=20
> Agreed its a extreme case, but illustrates the problem I have in mind
> for this patch.
>=20
> All tasks of same weight=3D1024.
>=20
> Before this patch
> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>=20
> sched_slice(grp A) =3D 20ms * 1/2 =3D 10ms
> sched_slice(A0) =3D 20ms
>=20
> sched_slice(grp B) =3D 20ms * 1/2 =3D 10ms
> sched_slice(B0) =3D (20ms * 1000/20) * 1 / 1000 =3D 1ms
> sched_slice(B1) =3D ... =3D sched_slice(B99) =3D 1 ms
>=20
> Fairness between groups and tasks would be obtained as below:
>=20
> A0 B0-B9 A0 B10-B19 A0 B20-B29
> |--------|--------|--------|--------|--------|--------|-----//--|=20
> 0 10ms 20ms 30ms 40ms 50ms 60ms
>=20
> After this patch
> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>=20
> sched_slice(grp A) =3D (20ms * 1001/20) * 1/2 ~=3D 500ms
> sched_slice(A0) =3D 500ms
Hmm, right that is indeed not intended
quoted text > sched_slice(grp B) =3D 500ms
> sched_slice(B0) =3D 0.5ms=20
This 0.5 is indeed correct, whereas the previous 1ms was not
quoted text > Fairness between groups and tasks would be obtained as below:
>=20
> A0 B0 - B99 A0
> |-----------------------|-----------------------|-----------------------=
|
quoted text > 0 500ms 1000ms 1500ms
>=20
> Did I get it right? If so, I don't like the fact that group A is allowed =
to run=20
quoted text > for a long time (500ms) before giving chance to group B.
Hmm, quite bad indeed.
quoted text > Can I know what real problem is being addressed by this change to
> sched_slice()?
sched_slice() is about lantecy, its intended purpose is to ensure each
task is ran exactly once during sched_period() - which is
sysctl_sched_latency when nr_running <=3D sysctl_sched_nr_latency, and
otherwise linearly scales latency.
