Introduce meaningful benchmarks

[qnikst]

We need better benchmarks, current ones are measuring very tiny action, and criterion does not provide enough precision for such checks. Additionally we miss some effects that may appear:

1. How metrics work under concurrency
2. How does metric affect program due to additional allocations (or no allocations)

Once we do that we need to clearly document it.