Abstract
Go is all about parallelism and concurrency, but they don’t come for free. This talk is about measuring their contention price, and being able to reduce it. A deep dive on contention: what is it, when is it expensive, and how to tackle it. Some light will be shed on ways to measure the computational power wasted waiting. The trace/block and pprof tools and libraries will be explained with actionable examples. After showing how to find contention issues, some handy ways to address them will be shown: •Splitting work: different ways to change common algorithms to not have a centralized state. •From channels to mutexes: where it makes sense to make the switch, and why. •From mutexes to atomics: the real price of making this step, and how to make an informed choice about it.
