Access to significant figures in benchmark timings?

[izaid]

Google Benchmark, while very useful, does seem to only provide timings in nanoseconds in the report. If I have a very tight function, which takes on the order of 2 nanoseconds, is it possible to get any of the significant figures here?

[EricWF]

No? I nanosecond is a pretty tiny amount of time. If you have a function that benchmarks at 2 nanoseconds it's likely the compiler optimized away your benchmark all together.

2 nanoseconds is like enough time to run 6 CPU instructions at most if I'm not mistaken.

[dmah42]

I don't think any of the methods we use for timing will get more precision. I doubt we get out of the way enough to allow 2ns to be a reasonable measure.

Can you force your benchmark to do the work N times per run and then divide the time measured by N?

[EricWF]

I did some tests and I'm 100% sure your benchmark is getting optimized away.

I tested an empty benchmark and a benchmark that only contained volatile int x = 42. The benchmarks ran in 1ns and 3ns respectively.

[EricWF]

@izaid I'm going to go ahead and close this issue based on what I think is happening. Do not hesitate to re-open it if you disagree. Sometimes I'm spectacularly wrong :-)

[izaid]

@EricWF Thanks, all of the suggestions here were good -- we can simply run N times the benchmark. I appreciate the feedback, you're right that I wasn't quite looking for the right thing.