* Serge Noiraud <serge.noiraud@bull.net> wrote:

> Hi,
> 
> I have a big dilemna on one machine :
> I run a task with RT priority which make a loop to mesure the system
> perturbation.
> It works well except on  one machine.
> On a multi-cpu, If I run the program on cpu 1, I get 23us. It's OK.
> If I run the same program on cpu 0, I get 17373us !
> If I do :
> # echo 0 >/proc/sys/kernel/preempt_max_latency

you can start/stop tracing from userspace too. Then you'll see what 
happens. See the attached trace-it.c code. After having activated 
user-triggered tracing, do something like this in your testcode:

	for (;;) {
	        gettimeofday(0, 1);
        	gettimeofday(0, 0);
	}

this should measure the maximum userspace delay on that CPU. You can 
also read the TSC in the inner loop to validate the kernel tracer:

	for (;;) {
	        gettimeofday(0, 1);
		t0 = rdtsc();
		t1 = rdtsc();
        	gettimeofday(0, 0);
		if (t1-t0 > max)
			print_max();
	}

(you should not measure the trace start/stop functions themselves, they 
can take alot of time when a maximum-latency event happens.)

	Ingo