Data Partitioning on Heterogeneous Multicore Platforms

Abstract

In this paper, we present two techniques for inter- and intra-node data partitioning aimed at load balancing MPI applications on heterogeneous multicore platforms. For load balancing between the multicore nodes of a heterogeneous multicore cluster, we propose how to define a functional performance model of an individual multicore node as a single computing unit, and use these models for data partitioning between the nodes. For load balancing within a heterogeneous multicore node, we propose a data partitioning technique between cores. Since parallel processes interfere with each other through shared memory, the speed of individual cores cannot be measured independently, and independent performance models cannot be defined for cores. Therefore, for a given problem size, we dynamically evaluate the performance of cores, while they are executing only the computational kernel of parallel application, and partition data proportionally to the observed speed.