Observatory benchmark utility library

This library contains various tools for observing the runtime behaviour of a program, in particular related to performance.

It is explicitly not intended to be a framework. That means unlike gbenchmark, etc., the user should not be expected to set up test benches in order to use this library.

Currently, three utilities have been implemented:

1. Perf Counters

2. Pyplot Sink and GGplot Sink

3. Perf Samplers

While these should work fine, the library was pretty much salvaged from a bunch of ad-hoc performance experiments and thus the level of polish is relatively low.

Design Notes

Compiler optimizations

Consider this example:

int foo() {
    ScopedTimer c(&result);

    int x = 0;
    for (int i=0; i<10000; ++i) x += i;
    return x;
}

The compiler is free to reorder this to

int foo() {
    ScopedTimer c(&result);
    c.~ScopedTimer();

    int x = 0;
    for (int i=0; i<10000; ++i) x += i;
    return x;
}

because there is no data dependency between measuring the time point and computing the return value. In my version of gcc this is done quite aggressively as soon as optimizations are turned on, and even marking result as volatile does not revert it. (Update April 2020: On my current version of gcc I have difficulties generating this effect at all, which probably kinda proves the point)

Because this issue is quite subtle to detect and can completely invalidate benchmarking results, most time measurements within observatory take an additional arbitrary parameter on which to generate a data dependency:

int foo() {
    ScopedTimer c(&result);
    int x = 0;
    for (int i=0; i<10000; ++i) x += i;
    c.split_after(x);
    return x;
}