To keep the cycle time of all perf benchmarks in check, we should set alert on benchmark_duration for all perf benchmarks (including ref build) with threshold of +5 minutes regressions.

I picked 5 minutes threshold without careful thought. We may need to tune the number a bit to make sure that benchmarking team is not overloaded with too many alerts.

+Annie: is there a way to try a specific alert threshold & see how many alerts would have been raised with the past data?