Topology-Aware Process Mapping on Clusters Featuring NUMA and Hierarchical Network
Dongyang Li, Yunlan Wang, Wei Zhu · 2013
Topology-aware process mapping plays a pivotal role in improving the performance of parallel programs on modern large-scale cluster computing systems featuring multicore architecture and hierarchical network. This paper firstly introduces a method to obtain application communication topology and hardware topology. Then a topology-aware process mapping algorithm named Isomorphic Tree Mapping is proposed for computing systems with hierarchical networks and NUMA effect. This algorithm partitions the sets of processes based on process proximity (or affinity) which is defined considering both the communication volume and frequency. During the partition of processes, a tree-like structure is built which is isomorphic to the physical topology tree. Then a mapping between any two correspondent nodes on the two isomorphic trees is performed. Experimental evaluations on a multicore cluster show that the process mapping method presented in this paper can effectively improve the communication efficiency of MPI application and cut down up to 10% of the whole execution time in most cases.