"Fast" in title. But I don't see any benchmarks, or discussion on how to reason about the performance of STM code.