Optimize latency data: remove the time when the benchmark generates data

[Supergirlzjl]

It is not reasonable that the latency data includes the time for benchmark generation. Can we add some optimization to remove the time for benchmark data generation from the latency data?