Towards Priority-Flexible Task Mapping for Heterogeneous Multi-core NUMA Systems

With the rapid development of heterogeneous multi-core processors, a new High Performance Computing (HPC) system architecture combining the heterogeneous multi-core architecture and NUMA architecture will emerge in the future. However, existing task mapping methods are ineffective on such systems because they do not simultaneously consider multiple performance factors caused by the heterogeneity in memory access and core performance. In a parallel application, one factor can affect performance more than another depending on the communication and computation load imbalances among parallel tasks. In this case, a task mapping method must prioritize one factor over another when calculating the mapping. To solve this problem, this paper proposes a new mapping method with two task mapping priority options: the memory-aware priority option (MPO) and the heterogeneity-aware priority option (HPO). A priority option switching mechanism (POSM) selects the appropriate priority option for the combination of a system and an application by analyzing their characteristics. Compared with other methods that do not switch mapping priorities, the proposed method achieves overall performance improvement when dealing with a set of applications with different characteristics.