bias results due to cpu "spin up"

[richemslie]

Thanks for making this public, I was about to do the same thing. :)

First few tests do suffer from "spin up" time of cpu ). Adding (some crappy code) to perform a dummy while loop for a second will knock about 20% time off the initial results.
eg:

int main(void) {
    auto start = std::chrono::steady_clock::now();
    long total = 0;
    while (true) {
        auto later = std::chrono::steady_clock::now();
        if (later - start > std::chrono::milliseconds(1000)) {
            break;
        }

        total++;
    }
    printf("Starting\n");
    ...

On i7-6900K, ubuntu 17.10

Without spinup

Performance of std::chrono::steady_clock
Average time per call: 20.1 ns
Clock tick period: 1.0 ns
Measured resolution: 18.0 ns (median: 20.0 ns) (sigma: 8.7 ns) (average: 20.1 +/- 0.1 ns)

Performance of std::chrono::system_clock
Average time per call: 20.3 ns
Clock tick period: 1.0 ns
Measured resolution: 19.0 ns (median: 20.0 ns) (sigma: 17.1 ns) (average: 20.3 +/- 0.1 ns)

Performance of std::chrono::high_resolution_clock
Average time per call: 19.1 ns
Clock tick period: 1.0 ns
Measured resolution: 14.0 ns (median: 20.0 ns) (sigma: 18.1 ns) (average: 19.1 +/- 0.1 ns)

With spinup:

Performance of std::chrono::steady_clock
Average time per call: 16.1 ns
Clock tick period: 1.0 ns
Measured resolution: 15.0 ns (median: 16.0 ns) (sigma: 6.3 ns) (average: 16.1 +/- 0.1 ns)

Performance of std::chrono::system_clock
Average time per call: 16.3 ns
Clock tick period: 1.0 ns
Measured resolution: 15.0 ns (median: 16.0 ns) (sigma: 8.2 ns) (average: 16.3 +/- 0.1 ns)

Performance of std::chrono::high_resolution_clock
Average time per call: 16.3 ns
Clock tick period: 1.0 ns
Measured resolution: 15.0 ns (median: 16.0 ns) (sigma: 9.7 ns) (average: 16.3 +/- 0.1 ns)