On Thu, Jun 19, 2008 at 12:27:14PM +0200, Peter Zijlstra wrote:
> On Thu, 2008-06-05 at 10:50 +0530, Ankita Garg wrote:
>
> > Thanks Peter for the explanation...
> >
> > I agree with the above and that is the reason why I did not see weird
> > values with cpu_time. But, run_delay still would suffer skews as the end
> > points for delta could be taken on different cpus due to migration (more
> > so on RT kernel due to the push-pull operations). With the below patch,
> > I could not reproduce the issue I had seen earlier. After every dequeue,
> > we take the delta and start wait measurements from zero when moved to a
> > different rq.
>
> OK, so task delay delay accounting is broken because it doesn't take
> migration into account.
>
> What you've done is make it symmetric wrt enqueue, and account it like
>
>   cpu0      cpu1
>
> enqueue
>  <wait-d1>
> dequeue
>             enqueue
>              <wait-d2>
>             run
>
> Where you add both d1 and d2 to the run_delay,.. right?
>

Thanks for reviewing the patch. The above is exactly what I have done.

> This seems like a good fix, however it looks like the patch will break
> compilation in !CONFIG_SCHEDSTATS && !CONFIG_TASK_DELAY_ACCT, of it
> failing to provide a stub for sched_info_dequeue() in that case.

Fixed. Pl. find the new patch below.
Signed-off-by: NAnkita Garg <ankita@in.ibm.com>
Acked-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Gregory Haskins <ghaskins@novell.com>
Cc: rostedt@goodmis.org
Cc: suresh.b.siddha@intel.com
Cc: aneesh.kumar@linux.vnet.ibm.com
Cc: dhaval@linux.vnet.ibm.com
Cc: vatsa@linux.vnet.ibm.com
Cc: David Bahi <DBahi@novell.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

We have the notion of tracking process-coupling (a.k.a. buddy-wake) via
the p->se.last_wake / p->se.avg_overlap facilities, but it is only used
for cfs to cfs interactions.  There is no reason why an rt to cfs
interaction cannot share in establishing a relationhip in a similar
manner.

Because PREEMPT_RT runs many kernel threads as FIFO priority, we often
times have heavy interaction between RT threads waking CFS applications.
This patch offers a substantial boost (50-60%+) in perfomance under those
circumstances.
Signed-off-by: NGregory Haskins <ghaskins@novell.com>
Cc: npiggin@suse.de
Cc: rostedt@goodmis.org
Acked-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Inspired by Peter Zijlstra.
Signed-off-by: NGregory Haskins <ghaskins@novell.com>
Cc: npiggin@suse.de
Cc: rostedt@goodmis.org
Acked-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

This patch fixes the following warning:

kernel/sched.c:1667: warning: 'cfs_rq_set_shares' defined but not used

This seems the correct way to fix this; cfs_rq_set_shares() is only used
in a single place, which is also inside #ifdef CONFIG_FAIR_GROUP_SCHED.
Signed-off-by: NVegard Nossum <vegard.nossum@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

fix:

kernel/sched.c: In function ‘sched_group_set_shares':
kernel/sched.c:8635: error: implicit declaration of function ‘cfs_rq_set_shares'
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Vegard Nossum reported:

> WARNING: at kernel/lockdep.c:2738 check_flags+0x142/0x160()

which happens due to:

 unsigned long long cpu_clock(int cpu)
 {
         unsigned long long clock;
         unsigned long flags;

         raw_local_irq_save(flags);

as lower level functions can take locks, we must not do that, use
proper lockdep-annotated irq save/restore.
Reported-by: NVegard Nossum <vegard.nossum@gmail.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

the rcutorture module relies on cpu_clock.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Signed-off-by: NDhaval Giani <dhaval@linux.vnet.ibm.com>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Measurement shows that the difference between cgroup:/ and cgroup:/foo
wake_affine() results is that the latter succeeds significantly more.

Therefore bias the calculations towards failing the test.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Increase the accuracy of the effective_load values.

Not only consider the current increment (as per the attempted wakeup), but
also consider the delta between when we last adjusted the shares and the
current situation.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

rw_i = {2, 4, 1, 0}
s_i = {2/7, 4/7, 1/7, 0}

wakeup on cpu0, weight=1

rw'_i = {3, 4, 1, 0}
s'_i = {3/8, 4/8, 1/8, 0}

s_0 = S * rw_0 / \Sum rw_j ->
  \Sum rw_j = S*rw_0/s_0 = 1*2*7/2 = 7 (correct)

s'_0 = S * (rw_0 + 1) / (\Sum rw_j + 1) =
       1 * (2+1) / (7+1) = 3/8 (correct

so we find that adding 1 to cpu0 gains 5/56 in weight
if say the other cpu were, cpu1, we'd also have to calculate its 4/56 loss
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

It was observed these mults can overflow.
Signed-off-by: NSrivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

We found that the affine wakeup code needs rather accurate load figures
to be effective. The trouble is that updating the load figures is fairly
expensive with group scheduling. Therefore ratelimit the updating.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

In case the domain is empty, pretend there is a single task on each cpu, so
that together with the boost logic we end up giving 1/n shares to each
cpu.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

The bias given by source/target_load functions can be very large, disable
it by default to get faster convergence.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

s_i = S * rw_i / \Sum_j rw_j

 -> \Sum_j rw_j = S * rw_i / s_i

 -> s'_i = S * (rw_i + w) / (\Sum_j rw_j + w)

delta s = s' - s = S * (rw + w) / ((S * rw / s) + w)
        = s * (S * (rw + w) / (S * rw + s * w) - 1)

 a = S*(rw+w), b = S*rw + s*w

delta s = s * (a-b) / b

IOW, trade one divide for two multiplies
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Priority looses much of its meaning in a hierarchical context. So don't
use it in balance decisions.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Currently task_h_load() computes the load of a task and uses that to either
subtract it from the total, or add to it.

However, removing or adding a task need not have any effect on the total load
at all. Imagine adding a task to a group that is local to one cpu - in that
case the total load of that cpu is unaffected.

So properly compute addition/removal:

 s_i = S * rw_i / \Sum_j rw_j
 s'_i = S * (rw_i + wl) / (\Sum_j rw_j + wg)

then s'_i - s_i gives the change in load.

Where s_i is the shares for cpu i, S the group weight, rw_i the runqueue weight
for that cpu, wl the weight we add (subtract) and wg the weight contribution to
the runqueue.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

doing the load balance will change cfs_rq->load.weight (that's the whole point)
but since that's part of the scale factor, we'll scale back with a different
amount.

Weight getting smaller would result in an inflated moved_load which causes
it to stop balancing too soon.
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

find_busiest_group() has some assumptions about task weight being in the
NICE_0_LOAD range. Hierarchical task groups break this assumption - fix this
by replacing it with the average task weight, which will adapt the situation.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

With hierarchical grouping we can't just compare task weight to rq weight - we
need to scale the weight appropriately.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Remove the fall-back to SCHED_LOAD_SCALE by remembering the previous value of
cpu_avg_load_per_task() - this is useful because of the hierarchical group
model in which task weight can be much smaller.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Finding the least idle cpu is more accurate when done with updated shares.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Re-compute the shares on newidle - so we can make a decision based on
recent data.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

While thinking about the previous patch - I realized that using per domain
aggregate load values in load_balance_fair() is wrong. We should use the
load value for that CPU.

By not needing per domain hierarchical load values we don't need to store
per domain aggregate shares, which greatly simplifies all the math.

It basically falls apart in two separate computations:
 - per domain update of the shares
 - per CPU update of the hierarchical load

Also get rid of the move_group_shares() stuff - just re-compute the shares
again after a successful load balance.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

We only need to know the task_weight of the busiest rq - nothing to do
if there are no tasks there.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

We used to try and contain the loss of 'shares' by playing arithmetic
games. Replace that by noticing that at the top sched_domain we'll
always have the full weight in shares to distribute.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

The idea was to balance groups until we've reached the global goal, however
Vatsa rightly pointed out that we might never reach that goal this way -
hence take out this logic.

[ the initial rationale for this 'feature' was to promote max concurrency
  within a group - it does not however affect fairness ]
Reported-by: NSrivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

It was observed that in __update_group_shares_cpu()

  rq_weight > aggregate()->rq_weight

This is caused by forks/wakeups in between the initial aggregate pass and
locking of the RQs for load balance. To avoid this situation partially re-do
the aggregation once we have the RQs locked (which avoids new tasks from
appearing).
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Keeping the aggregate on the first cpu of the sched domain has two problems:
 - it could collide between different sched domains on different cpus
 - it could slow things down because of the remote accesses
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

show all the schedstats in /debug/sched_debug as well.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Uncouple buddy selection from wakeup granularity.

The initial idea was that buddies could run ahead as far as a normal task
can - do this by measuring a pair 'slice' just as we do for a normal task.

This means we can drop the wakeup_granularity back to 5ms.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

with sched_clock_cpu() being reasonably in sync between cpus (max 1 jiffy
difference) use this to provide cpu_clock().
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Try again..

Initial commit: 18d95a28
Revert: 6363ca57Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Ok, so why are we in this mess, it was:

  1/w

but now we mixed that rw in the mix like:

 rw/w

rw being \Sum w suggests: fiddling w, we should also fiddle rw, humm?
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

calc_delta_asym() is supposed to do the same as calc_delta_fair() except
linearly shrink the result for negative nice processes - this causes them
to have a smaller preemption threshold so that they are more easily preempted.

The problem is that for task groups se->load.weight is the per cpu share of
the actual task group weight; take that into account.

Also provide a debug switch to disable the asymmetry (which I still don't
like - but it does greatly benefit some workloads)

This would explain the interactivity issues reported against group scheduling.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Try again..

initial commit: 8f1bc385
revert: f9305d4aSigned-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

In file included from /mnt/build/linux-2.6/kernel/sched.c:1496:
/mnt/build/linux-2.6/kernel/sched_rt.c: In function '__enable_runtime':
/mnt/build/linux-2.6/kernel/sched_rt.c:339: warning: unused variable 'rd'
/mnt/build/linux-2.6/kernel/sched_rt.c: In function 'requeue_rt_entity':
/mnt/build/linux-2.6/kernel/sched_rt.c:692: warning: unused variable 'queue'
Signed-off-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Conflicts:

	kernel/sched_rt.c
Signed-off-by: NIngo Molnar <mingo@elte.hu>