Analysis and optimization of CFS scheduler on NUMA-based systems

Non Uniform Memory Access (NUMA) architecture becomes more and more popular as it has better scalability than Uniform Memory Access (UMA). However, all previous work on the operation system scheduler assumed that the underlying system is UMA. As a result, the performance degrades on NUMA machines due to lacking of consideration to the underlying hardware. Researchers discover that the Completely Fair Scheduler (CFS) does not work smoothly on NUMA machines and even interfere performance relative to the O (1) scheduler. In this paper researchers investigate the causes for the performance decline and devise an architecture aware task-bound approach for NUMA system, which can help the CFS scheduler works efficiently on NUMA platforms. The evaluation shows that the approach can upgrade the system performance by more than 60 % on average. The research has great significance to the development and popularity of domestic operating system.