On the use of communication graphs in automatic mapping

The assignment of processes to processors (the mapping problem) is one of the major factors affecting the performance of a parallel application. Since the problem of finding an optimal mapping is known to be NP-complete many heuristic strategies have been devised to obtain “good” mappings. Effective parallel programming environments therefore include a mapper, a tool that, starting from a model of the parallel application, maps processes to processors without the user intervention. The model used by most tools is the communication graph. The authors discuss the role of communication graphs as models for parallel application and show some weak points of the approach even in those situations that are usually considered most favourable for the use of communication graphs, namely parallel applications whose run-time behaviour can be determined through a static analysis