On Wed, Jun 04, 2014 at 10:32:10AM +0100, Vincent Guittot wrote:

On 4 June 2014 10:08, Peter Zijlstra [off-list ref] wrote:

quoted

On Wed, Jun 04, 2014 at 09:47:26AM +0200, Vincent Guittot wrote:

quoted

On 3 June 2014 17:50, Peter Zijlstra [off-list ref] wrote:

quoted

On Wed, May 28, 2014 at 04:47:03PM +0100, Morten Rasmussen wrote:

quoted

Since we may do periodic load-balance every 10 ms or so, we will perform
a number of load-balances where runnable_avg_sum will mostly be
reflecting the state of the world before a change (new task queued or
moved a task to a different cpu). If you had have two tasks continuously
on one cpu and your other cpu is idle, and you move one of the tasks to
the other cpu, runnable_avg_sum will remain unchanged, 47742, on the
first cpu while it starts from 0 on the other one. 10 ms later it will
have increased a bit, 32 ms later it will be 47742/2, and 345 ms later
it reaches 47742. In the mean time the cpu doesn't appear fully utilized
and we might decide to put more tasks on it because we don't know if
runnable_avg_sum represents a partially utilized cpu (for example a 50%
task) or if it will continue to rise and eventually get to 47742.

Ah, no, since we track per task, and update the per-cpu ones when we
migrate tasks, the per-cpu values should be instantly updated.

If we were to increase per task storage, we might as well also track
running_avg not only runnable_avg.

I agree that the removed running_avg should give more useful
information about the the load of a CPU.

The main issue with running_avg is that it's disturbed by other tasks
(as point out previously). As a typical example,  if we have 2 tasks
with a load of 25% on 1 CPU, the unweighted runnable_load_avg will be
in the range of [100% - 50%] depending of the parallelism of the
runtime of the tasks whereas the reality is 50% and the use of
running_avg will return this value

I'm not sure I see how 100% is possible, but yes I agree that runnable
can indeed be inflated due to this queueing effect.

In fact, it can be even worse than that because i forgot to take into
account the geometric series effect which implies that it depends of
the runtime (idletime) of the task

Take 3 examples:

2 tasks that need to run 10ms  simultaneously each 40ms. If they share
the same CPU, they will be on the runqueue 20ms (in fact a bit less
for one of them), Their load (runnable_avg_sum/runnable_avg_period)
will be 33% each so the unweighted runnable_load_avg of the CPU will
be 66%

Right, it actually depends on how often you switch between the tasks. If
you sched_tick happens every 10 ms then in this example one task will
run for 10 ms an be done, while the other one waits for 10 ms and then
runs to completion in 10 ms. The result is that one task is runnable for
10 ms and the other one is runnable for 20 ms. That gives you 25% and
50% for a total of 75%.

2 tasks that need to run 25ms simultaneously each 100ms. If they share
the same CPU, they will be on the runqueue 50ms (in fact a bit less
for one of them), Their load (runnable_avg_sum/runnable_avg_period)
will be 74% each so the unweighted runnable_load_avg of the CPU will
be 148%

2 tasks that need to run 50ms  simultaneously each 200ms. If they
share the same CPU, they will be on the runqueue 100ms (in fact a bit
less for one of them), Their load
(runnable_avg_sum/runnable_avg_period) will be 89% each so the
unweighted runnable_load_avg of the CPU will be 180%

In this case you switch been the tasks before they complete, so in this
case both tasks are waiting so runnable will get much righer than 75%.
You are right. It was thinking about tasks where the busy period is
short enough that they run to completion when they are scheduled.