benchmark.cc is using non portable high precision timer

[ruipires]

Currently benchmark.cc is calling clock_gettime with CLOCK_PROCESS_CPUTIME_ID as the first parameter.
This is not available in Mac OSX (or windows AFAIK), and results in failure to compile.

I'm considering sending a pull request, but am not sure how exactly to fix this.

Using #ifdefs to handle different platforms is something I tend to avoid at all costs.
Considering c++11 is an option here, using std::chrono::high_resolution_timer could help, but that should measure wall time and not cpu time.
std::clock could be a cross platform way to implement this, but seems to have less resolution than clock_gettime with CLOCK_PROCESS_CPUTIME_ID.

[craigbarnes]

Wouldn't clock() / CLOCKS_PER_SEC be good enough for this kind of benchmark? It's by far the most portable way to measure duration (it's C89) and it's quirks only apply when timing I/O, which is all done before the timer starts in benchmark.cc.

[nostrademons]

I like the clock() / CLOCKS_PER_SEC approach. The reason I didn't use it originally was:

1. I was worried about picking up variance through I/O, yielding to other processes, etc.
2. I wanted higher resolution than millisecond.

But reading the spec for clock(), it seems like it's processor time for this process only, and on modern systems gives microsecond resolution. And considering that I can't seem to get numbers more accurate than about +-200ms on my laptop because of Turbo Boost, I think that the desire for more accuracy may not be worth it, particularly at the cost of portability.

[nostrademons]

Reproduced on Travis CI, the pull request fixed it.

[ruipires]

👍