Exploiting Geometric Partitioning in Task Mapping for Parallel Computers

Abstract—We present a new method for mapping applica-tions ’ MPI tasks to cores of a parallel computer such that communication and execution time are reduced. We consider the case of sparse node allocation within a parallel machine, where the nodes assigned to a job are not necessarily located within a contiguous block nor within close proximity to each other in the network. The goal is to assign tasks to cores so that interdependent tasks are performed by “nearby ” cores, thus lowering the distance messages must travel, the amount of con-gestion in the network, and the overall cost of communication. Our new method applies a geometric partitioning algorithm to both the tasks and the processors, and assigns task parts to the corresponding processor parts. We show that, for the structured finite difference mini-app MiniGhost, our mapping method reduced execution time 34 % on average on 65,536 cores of a Cray XE6. In a molecular dynamics mini-app, MiniMD, our mapping method reduced communication time by 26% on average on 6144 cores. We also compare our mapping with graph-based mappings from the LibTopoMap library and show that our mappings reduced the communication time on average by 15 % in MiniGhost and 10 % in MiniMD. I