Currently, the harness uses larger strides to run and consume the results of the test. Unfortunately, when actor threads are unbalanced (e.g. run asymmetric workloads), there is a single collision per entire stride. Having shorter strides would increase collision ratios, but would create much more overhead for creating/consuming benchmark results and state.

We can do long "epoch" with several shorter "strides" and spin-loop checkpointing between strides, so that threads can catch up more often.