Eecient Program Partitioning Based on Compiler Controlled Communication 1

In this paper, we present an e cient framework for intraprocedural performance based program partitioning for sequential loop nests. Due to the limitations of static dependence analysis especially in the inter-procedural sense, many loop nests are identi ed as sequential but available task parallelism amongst them could be potentially exploited. Since this available parallelism is quite limited, performance based program analysis and partitioning which carefully analyzes the interaction between the loop nests and the underlying architectural characteristics must be undertaken to e ectively use this parallelism. We propose a compiler driven approach that con gures underlying architecture to support a given communication mechanism. We then devise an iterative program partitioning algorithm that generates e cient program partitioning by analyzing interaction between e ective cost of communication and the corresponding partitions. We model this problem as one of partitioning a directed acyclic task graph (DAG) in which each node is identi ed with a sequential loop nest and the edges denote the precedences and communication between the nodes corresponding to data transfer between loop nests. We introduce the concept of behavioral edges between edges and nodes in the task graph for capturing the interactions between computation and communication through parametric functions. We present an e cient iterative partitioning algorithm using the behavioral edge augmented PDG to incrementally compute and improve the schedule. A signi cant performance improvement (factor of 10 in many cases) is demonstrated by using our framework on some applications which exhibit this type of parallelism.