Data and Process Alignment in Modula-2*
Michæl Philippsen, Markus U. Mock · 1994
Exploiting locality is a central goal of translating problem-oriented parallel programming languages for distributed memory parallel machines. Modula-2* places the burden of automatically deriving good data and process distribution on the compiler. In this paper we present a technique implemented in our optimizing compiler that enhances locality in a source-to-source transformation. Analysis of data access patterns and parallel operations leads to an arrangement graph. Processing of this graph reveals conflicting arrangements. Some assumptions and a heuristic based on dynamic programming enables the compiler to find the best alignment in logarithmic time. The technique has improved runtime performance on benchmarks by over 60%. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.