Use the memory profiler only for the memory benchmarks. This reduces the time per iteration, allowing to do more iterations. This also gives a more natural setup, without explicit gc calls between iterations. All this reduces the variance of the results.