Re: [patch] Real-Time Preemption, -RT-2.6.10-rc2-mm3-V0.7.32-15

[Mark_H_Johnson]

> * Mark_H_Johnson@xxxxxxxxxxxx <Mark_H_Johnson@xxxxxxxxxxxx> wrote:
>
> > The maximum duration of the CPU loop (as measured by the application)
> > is in the range of 1.42 msec to 2.57 compared to the nominal 1.16 msec
> > duration for -20RT.  The equivalent numbers for -20PK are 1.28 to 1.93
> > msec. [...]
>
> so -20RT has resolved all the CPU-loop-max-delay issues of the -RT
> kernel regarding the RT-priority CPU loop and in essence adds only a
> small amount of delay (100 usecs?) to the nominal (==minimum possible)
> delay?
I believe so, or in other words a minor penalty to average latency impact
to get a reduction in maximum latency [though I am not sure I actually
measured such a reduction].

> i suspect the 100 usecs comparison is an effect of the cutoff value
> being a single value. Also, 100 usecs is so close to the DMA related
> delay which makes it hard to compare it - other than stating that -RT
> has higher CPU overhead.
I think we agree on this assessment. I don't think the threading overhead
is 100 usec but more likely 50 usec [at least on my hardware - see below].
That the extra 50 usec puts us over the 100 usec measure is unfortunate.

> are the ping times still considered anomalous? Could be a side-effect of
> the different flow of control between hardirq/softirq contexts. (There
> have been a (low but nonzero) number of assumptions about the flow in
> pieces of softirq code, and there could be more.)
Please note that to get the ping times I stated, I set the priority of
ksoftirqd/0 and /1 to 99 (along with the IRQ threads). Others won't be
so generous and get different results.

I believe the longer duration with -20RT is due to the threading overhead.
Look at the numbers for -20PK and -20RT, the minimums are
  -20PK 0.089 msec (or 089 usec)
  -20RT 0.134 msec (or 134 usec)
or a difference of about 50 usec. If the sequence of steps is something
like
  Network Interrupt -T-> Network IRQ -T-> ksoftirqd/0 -> ping reply
on -RT and
  Network Interrupt -> Network IRQ -> soft IRQ -> ping reply
on -PK, the difference measured is likely reasonable.

[where -T-> represents a new thread activation, and -> represents something
like a function call]

Again, the benefit of PREEMPT_RT (over PREEMPT_DESKTOP) should be reduced
maximum latencies with a modest (or very little) impact to overall
performance / throughput. We are certainly getting a lot closer on
reaching that goal.

--Mark H Johnson
  <mailto:Mark_H_Johnson@xxxxxxxxxxxx>

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/