3

I have a computer program.

I profile it with a test bench like:

for (size_t ii = 0U; ii < LOOP; ++ii) {
    clock_gettime(CLOCK_MONOTONIC, &time_start[ii]);
    do_work();
    clock_gettime(CLOCK_MONOTONIC, &time_finish[ii]);
}

However, for microbenchmarking purposes clock_gettime starts to interfere and the x86 rdtsc counter has oddities that make it annoying to use.

Is it valid to instead profile:

for (size_t ii = 0U; ii < LOOP; ++ii) {
    clock_gettime(CLOCK_MONOTONIC, &time_start[ii]);
    for (size_t jj = 0U; jj < INNER_LOOP; ++jj) {
        do_work();
    }
    clock_gettime(CLOCK_MONOTONIC, &time_finish[ii]);
}

and average the time taken using INNER_LOOP?

I have a long-tailed distribution that looks like a Cauchy distribution in the beginning.

Long tail distribution

The actual latencies involved are really large at the deep end though.

really long tail distribution

The distribution does not seem to follow a power law.

It seems to me that taking the average is inappropriate for a long-tail distribution of my sort. I might need to take the median, maximum, minimum or some other quantity instead. I am not sure of the maths behind why this should be so though and what my alternative are.

Stephan Kolassa
  • 95,027
  • 13
  • 197
  • 357
Molossus Spondee
  • 481
  • 3
  • 9
  • 1
    I do think there's an on-topic question in here if you could focus it on the statistical analysis in some way, but at the moment it isn't clear what kind of answer you are after. – Silverfish Nov 21 '17 at 23:07
  • There is some partial answer in https://stats.stackexchange.com/questions/94402/what-is-the-difference-between-finite-and-infinite-variance/100161#100161 – kjetil b halvorsen Aug 10 '18 at 09:00

0 Answers0