Require benchmarks to make performance improvements

[joestringer]

If someone is designing a CFP which optimizes performance, it might be a useful datapoint to have benchmarks in the CFP template. This would allow us to identify assumptions in the performance optimizations and help to guide the design based on real-world data.

[bowei]

+1 on this, this guideline will prompt contributors to do something that usually comes out in the discussions.

[msune]

If someone is designing a CFP which optimizes performance, it might be a useful datapoint to have benchmarks in the CFP template.

I would suggest to have a set of benchmark tests in Cilium to run against as a proof of improvement (I think we have some of that).

It's very easy to have incorrect / biased measurements, as more often than not setup, configuration, interference with other SW, HW etc. - even in relative comparisons - play a major role in the measurements.

This should also help spotting regressions on non-performance patches. To me this is a bigger concern than performance patches themselves.