hodgesds / perf-utils Goto Github PK

I would like to use your library to count cycles and instructions as a part of a standard Go benchmark for one of my projects. Unfortunately, the current work flow seems to be not as useful as it could be for this purpose: the profileFn function performs the thread locking and set up in one go, causing all the cost for this to become part of the allocation and time counts measured by the benchmark. Additionally, the current set up only allows taking one performance counter per benchmark when taking multiple counters would be a lot more useful.

I think it would be nice to have a function with a signature like this:

func RunBenchmark(b *testing.B, f func(b *testing.B), unix.PerfEventAttr...)

This function would do the following:

• lock the current Go routine to a thread and set scheduler affinity
• call b.StopTimer(); b.ResetTimer()
• run f once for each PerfEventAttr and add the results (divided by b.N) to the benchmark using b.ReportMetric
• run f once more surrounded by b.StartTimer() and b.StopTimer to get timings and GC usage (optionally do this first?)
• if a performance counter cannot be measured, it is silently omitted (and optionally logged?)
• other errors are reported using b.Fatal()

What do you think about something like this? Alternatively or additionally, a benchmark setup function (using b.Cleanup to automatically clean up) and a setup-less profileFn variant could be provided.