Hi Sebastian

As you got correctly i was talking about 4.4.1-rt5 and not 4.1 i replied to by 
accident.

Am Freitag, 12. Februar 2016, 10:07:59 schrieb Sebastian Andrzej Siewior:

On 02/12/2016 09:28 AM, Tim Sander wrote:

quoted

Hi Sebastian

Hi Tim,

quoted

Am Sonntag, 16. August 2015, 15:56:30 schrieb Sebastian Andrzej Siewior:

quoted

I'm pleased to announce the v4.1.5-rt5 patch set.

I have just tested it with a Altera SoC ARM v7. The latencies seem to have
gotten a little bit worse with each release. The first core has always
been
worse (presumably due to interrupt load) but now it dropped to 111µs (rt5)
from 76µs(rt3) and 54µs(rt2).

in -rt2 we had bug in migrate disable code which means each task was
running on CPU0. This got partly fixed in -rt3. In -rt3 the scheduler
could assign a task to CPU1 but the task should stay there for ever.
This little detail was fixed in -rt5.
This is one thing that comes to mind.
Lazy-preempt should have been fixed in -rt3, too. This should not give
you higher latencies but higher throughput.

What about rt4? It is only the stable update so you should see here the
numbers from rt3. If that is true and your numbers are stable it should
be easy to run git bisect between rt4 and rt5. And looking at
  https://git.kernel.org/rt/linux-rt-devel/h/v4.4.1-rt5
the only non-cosmetic change in -rt5 that should affect you is the
migrate-disable fixup from Mike.

Ok, each run takes a couple of hours so bisecting should take quite some time
but i will give it a try. I started a test with 4.4.1-rt4, if the numbers are 
within the 70µs ballpark bisecting seems the way to go. If the numbers are 
higher i suspect that stable update might have a play here. But we will see.

Best regards
Tim