John,

On Sat, May 14, 2011 at 6:28 PM, Thomas Gleixner[off-list ref]  wrote:

quoted

On Sat, 14 May 2011, John Kacur wrote:

quoted

I did some light testing merging 2.6.33.13 into real-time 2.6.33.9-rt31.
In addition I cherry-picked 3c955b407a084810f57260d61548cc92c14bc627
in order to compile on newer distros.

Thanks.

quoted

Here is the result of cyclic test on one machine
sudo ./cyclictest -t32 -p 80 -n -i 10000 -l 10000
policy: fifo: loadavg: 0.00 0.00 0.00 1/541 3759

T: 0 ( 3728) P:80 I:10000 C:  10000 Min:      7 Act:  104 Avg:  114 Max:     470

The numbers are weird. How does that compare to older kernels on that
machine with the same test?

Note this was on a machine with a straight Fedora install, and no -rt
packages or tuning. That being the case, the numbers are no better or
worse than recent -rt kernels. I can get you numbers with the last
kernel if you wish, on Monday. Perhaps Carsten would be interested in
running his tests?

I created my own 2.6.33.13-rt31 kernel some days ago. It is running on a 
pretest machine in our farm (rack #1, slot #4) since then -> 
https://www.osadl.org/?id=994. For an overview about the kernels in the 
farm (ordered by kernel version and release) see -> 
https://www.osadl.org/?id=1001. Real-time and all other data before and 
after the kernel upgrade to 2.6.33.13-rt31 are indistinguishable between 
each other.

Nevertheless, it would be interesting to investigate the origin of the 
numbers Thomas (correctly IMHO) classified as weird. The average and the 
maximum latency could be from a Pentium 90, the minimum latency looks ok 
for a state-of-the-art processor. What machine is this and how did you 
manage to create such numbers? It doesn't look like SMIs were 
responsible for that since they normally increase the maximum, not the 
average. Was another task with prio 80 or higher running?

Thanks,
Carsten.