Efficient grouping-based mapping and scheduling on heterogeneous cluster architectures

Heterogeneous clusters of computers usually provide high computing power for large-scale applications at the expense of large cost. And there are two challenges currently faced by researchers. One is how to map large applications, modeled by Directed Acyclic Graphs (DAG), to heterogeneous architectures with minimal cost. The other is how to schedule tasks on each cluster to further decrease the total cost. This paper proposes an Integer Linear Programming (ILP) formulation to achieve optimal results. In order to improve the efficiency for ILP, the problem size is first reduced by Safe Graph Grouping (SGG). SGG guarantees the reduced graph to be a DAG to avoid dead lock during the scheduling. Then Time Constrained Local Scheduling (TCLS) is used to further reduce total cost for the input task graph. Experimental results show that the cooperation of SGG and ILP can reduce the total cost by 11.46% compared with existing mapping techniques.