Mapping Heterogeneous Task Graphs onto Networks: Execution Time Optimization

Abstract

The paper presents a formulation for the problem of mapping parallel programs on heterogeneous networks and proposes a distributed recursive (heuristic) algorithm for its solution. This algorithm doesn’t require global knowledge of computational state, it uses only information obtained from a neighbours nodes. Parallel programs and networks are presented as weighted graphs. In each stage graph bisection strategy is used to divide all processes of the program on two groups, which farther may be sent to neighbour nodes, where algorithm will be continued in the same way, or be leaved on initial node.