High performance computing (HPC) architectures are specialized machines which can reach their peak performance only if they are programmed in a way which exploits the idiosyncrasies of the architecture. An important feature of most such architectures is a physically distributed memory, resulting in the requirement to take data locality into account independent of the memory model offered to the user. In this paper we discuss various ways for managing data distribution in a program, comparing in particular the low-level message-passing approach to that in High Performance Fortran (HPF) and other high performance languages. The main part of the paper outlines a method for the specification of data distribution semantics for distributed-memory architectures and clusters of SMPs. The paper concludes with a discussion of open issues and references to future work.

parallel architectures, distributed memory, high performance computing, data distribution