Euphrates: A System for Automatic Introduction of Data Parallelism into Modular Applications Andrew J.S. Wilson lan W. Flockhart The University of Edinburgh migrating an application from the MAB to a stand-alone program suitable for execution on a serial or parallel computer. Topological analysis of a network of modules allows derivation of the order In which each processing module must be applied to the input data set. Analysis of the data access patterns of each module allows estimation of memory requirements. This allows large data sees to be processed in a tile-by-tile fashion, on computers with relatively small amounts of memory. The parallelizatJon by data decomposition of a wide class of useful processing operations is well understood and supported by a variety of parallel utility libraries. By rigorously defining the classes of operations supported, the Euphrates system is able to automatically generate parallel applications which will process arbitrarily large data sets, a capability which we believe to be unique [7]. These applications are parallelized using regular geometric data decomposition and are demonstrated to perform correctly and at increasing speed with increasing machine size. Future work will concentrate on widening the base of implemented modules and the class of supported
/lp/association-for-computing-machinery/euphrates-a-system-for-automatic-introduction-of-data-parallelism-into-7M790gfNC1