An incremental methodology for parallelizing legacy stencil codes on message-passing computers

There is considerable interest in converting existing production sequential/vector codes into scalable portable parallel message-passing programs. However the task is very tedious and error prone. In this paper we describe a systematic incremental methodology that greatly eases parallelization of stencil codes, by operating in a single address space while addressing data partitioning issues and then transitioning to multiple address spaces to handle communication and synchronization issues. The approach has been successfully used in manually parallelizing a production CFD code. Since the steps in the methodology are formalized as algorithms, the approach is amenable to automation