Automatic parallelization tools and their applications

Abstract

In parallelizing huge legacy codes such as NCARIPenn State MMS, a proper software environment is critical for reducing the time and effort. This thesis presents an empirical study of automatic parallelization based on the NCAR/Penn State MMS model, the Pacific Northwest National Laboratory (PNNL) version of MMS and a FDM benchmark program. ParAgent, a tool for automatic parallelization, VisS-D a visualization tool, a web-based monitor, and Rabbit, a performance analysis tool were used in this study. In addition, ahigh-level communication library was developed to complement the use of ParAgent. Performance is one of the most important aspects of parallelism. We tested different types of networks and PC clusters to see how the communication between processors affects performance. Also, we put some efforts in analyzing and reducing load imbalance.