Ingo Molnar wrote:
> i have released the -V0.7.30-9 Real-Time Preemption patch, which can be
> downloaded from the usual place:
> 

A couple of observations that I would like to share:

1) With all of the tracing compiled in, there is a measurable amount of 
additional latency, even when the tracing is disabled via the proc 
interfaces. So if you want to measure best-case (or worst-case) 
latencies turn off the tracing facilities to do so. For example, using a 
blocking read, the worst case latency from interrupt to read: wo/tracing 
~68 usec, w/tracing ~110 usec. I am attaching two histograms generated 
by the rtc histogram by running a slightly modified version of realfeel.

2) Also there is a big difference in latencies between blocking reads 
and using signals to notify the process that data is available. I am 
also attaching a histogram generated by the rtc histogram by running 
amlat. Part of the numbers in the > 10 msec bucket are into the hundreds 
of msecs.

All of these numbers were generated using a comparable load from 
stress-kernel.

kr