PREDICTIVE CPU ISOLATION OF CONTAINERS AT NETFLIX

Summary from the TLDR newsletter

Computer architecture design has evolved towards adding various levels of caching between compute units and the main memory to hide latency. These caches are partially shared among the CPUs, so perfect performance isolation of co-hosted containers is not possible. Netflix's team was able to meaningfully improve both the predictability and performance of its containers by taking some of the CPU isolation responsibility away from the operating system and moving towards a data-driven solution involving combinatorial optimization and machine learning. This post explains the process in detail.