Abstract: From the perspective of data redistribution, traditional MPI automatic parallelizing compilation systems gen- erate message-passing programs for distributed-memory systems,but a large number of data redistribution communica- lion overheads result in their low speedups. Aiming at the problem, this paper proposed a message-passing code genera- lion algorithm in the back-end of the MPI automatic parallelizing compilation system based on Open64. With the centre of uniform data distribution, the algorithm generates more accurate message-passing codes, according to the given sets of parallel loops and communication arrays, by modifying the WHIRL syntax trees of serial codes. Experimental results show that the algorithm can reduce communication overheads of messagcpassing programs to a large extent and im- prove their speedups significantly.

Key words: MPI, Automatic parallelizing compilation, Distributed-memory system, Messagcpassing code, Open64,Speedup