Martijn van Oosterhout <kleptog@svana.org> writes:
> Wierd. Can you get the output of *instr in each call of
> InstrEndLoop? Preferably after it does the calculation but before it
> clears the values. So we get an idea of number of samples and what it
> guesses. SampleOverhead would be good too.
The problem looks to be an underestimation of SampleOverhead, and on
reflection it's clear why: what CalculateSampleOverhead is measuring
isn't the total overhead, but the time between the two gettimeofday
calls. Which is probably about half the true overhead. What we
ought to do is iterate InstStartNode/InstrStopNode N times, and
*separately* measure the total elapsed time spent.
It occurs to me that what we really want to know is not so much the
total time spent in InstStartNode/InstrStopNode, as the difference in
the time spent when sampling is on vs when it is off. I'm not quite
sure if the time spent when it's off is negligible. Off to do some
measuring ...
regards, tom lane
