A Parallelizing Compiler for the Workstaion Cluster Environment

Abstract

In this paper, we propose a coarse grain process generation method suitable for the workstation cluster environment. In our method, we adapt the functional parallelizing strategy, in which a function written in a given program is treated as a unit of process generation. Namely, our process generation algorithm decides whether each function is executed as an independent distributed process or as a normal function call. For this decision, the dependency value is calculated to evaluate the parallel executability between callee/caller functions. Our process generation method can generate the processes with coarser granularity and lower communication frequency. The workstation cluster environment consists of several autonomous workstations connected through the local-area-network, and has no parallel-processing-specific facilities. Thus, the overheads of process invocation and communication among processes are very large in comparison with the parallel computer environment which other methods deal with. Therefore, considering the property of our processes described above, our process generation method is suitable for the distributed computer environment in term with the process granularity and the communication frequency.