Re: [patch] Real-Time Preemption, -RT-2.6.10-rc1-mm3-V0.7.23

[Mark_H_Johnson@raytheon.com]

>i have released the -V0.7.23 Real-Time Preemption patch, which can be
>downloaded from the usual place:
>
>    http://redhat.com/~mingo/realtime-preempt/

A few notes on results from running with this patch (well, actually -EA
that Ingo provided separately).

[1] Build problems, separately reported to Ingo with CONFIG_PREEMPT_RT
enabled on x86 and you have modules using kunmap_atomic. Fix by adding
kunmap_virt and kmap_to_page to the list of exports.

[2] The live lock that I was having seems to have been killed based
on an hour of testing (I could usually cause it in 5 minutes or less).

[3] I am not so sure that the latency tracing works. I do not get any
trace output, even if I set preempt_max_latency to zero. I also noticed
that /proc/sys/kernel/preempt_wakeup_timing was removed at .20 but
not sure if that was deliberate. As a result, I have no reports from
the kernel tracing.

[4] Application level latencies are OK but not great.
 X test - only 90% of CPU loops are within 100 usec of nominal value.
In previous RT kernels I got > 99% with 100 usec.
 top test - looks much nicer than X test, but still have up to 30%
overhead on CPU loop.
 network I/O tests - smoothest of all test results, very nice
 disk write - very noisy, bursts of long delays with only 82% within
100 usec and worst case has over 100% overhead (2.5 msec vs 1.16 nominal)
 disk copy - fewer bursts, but worst case is similar to disk write.
About 95% within 100 usec.
 disk read - relatively clean with a pair of modest bursts early in
the test and then settled out a little worse than the network tests.
99.9% of samples within 100 usec and max of 1.65 msec.

[5] concurrent ping of system had over 13% lost packets (1089 out
of 10723 - plus I let this run after the tests had completed). The
2.4 RT kernel I use has no lost packets.

[6] I did a separate run of a script Ingo suggested that samples
the kernel profile data. It shows basically no CPU time for the
events/0 and events/1 tasks. I took 11 samples in about 1/2 hour
of testing but nothing seems to jump out of the data.

--Mark H Johnson
  <mailto:Mark_H_Johnson@raytheon.com>