NUMA-Aware Contention Scheduling on Multicore Systems

Abstract

More and more modern multi-processor systems have adopted Non-uniform Memory Access (NUMA) architecture, and each processor uses a chip multiprocessor (CMP) architecture. Therefore, some hardware resources (such as the last-level cache and memory controller) are shared among multiple physical cores in a single processor. However, contention for shared resources may cause severe and unpredictable performance impacts on threads running on the CMP. Previous work has studied alleviating resource contention by scheduling threads or migrating memory. However, because NUMA has multiple nodes, nodes with serious contention for shared resources may change at any time. Simply scheduling threads or migrating memory may cause the second competition for shared resources. After analyzing the memory access control flow under the NUMA architecture, this paper proposes a competition-aware scheduling algorithm, and designs a NUMA competition-aware framework based on PMU and eBPF, which dynamically perceives the shared resource competition load of NUMA nodes (hot nodes and cold nodes), and adjust the mapping relationship between threads and NUMA nodes in memory according to the contention load.