Hierarchical task mapping for parallel applications on supercomputers

As the scale of supercomputers grows, so does the size of the interconnect network. Topology-aware task mapping, which maps parallel application processes onto processors to reduce communication cost, becomes increasingly important. Previous works mainly focus on the task mapping between compute nodes (i.e., inter-node mapping), while ignoring the mapping within a node (i.e., intra-node mapping). In this paper, we propose a hierarchical task mapping strategy, which performs both inter-node and intra-node mapping. We consider supercomputers with popular fat-tree and torus network topologies, and introduce two mapping algorithms: (1) a generic recursive tree mapping algorithm, which can handle both inter-node mapping and intra-node mapping; (2) a recursive bipartitioning mapping algorithm for torus topology, which efficiently partitions the compute nodes according to their coordinates. Moreover, a hierarchical task mapping library is developed. Experimental results show that the proposed approach significantly improves the communication performance by up to 77 % with low runtime overhead.