by
Will Denissen1, Vincent Korstanje1, Peter Maarleveld1 and Henk J. Sips2
1 TPD-TNO, Delft, the Netherlands
2 Delft University of Technology, Delft, The Netherlands
Abstract
Converting applications to HPF is a major investment for industrial users. The conversion process has many steps and many decision points. To facilitate the converting process and the users willingness to proceed, two things are of importance: (1) an estimation of the amount of (re)coding work for the application at hand and (2) the availability of a methodology and supporting tools to support the process of application analysis and (re)coding.
At the HPFC knowledge center in Delft, the Netherlands, we have developed a methodology that yields an estimate for the amount of migration work, the possible speed up that can be expected, an identification of the most time-consuming parts, and a classification of the type of parallelism needed for the application. The latter is important for codes which cannot be solely migrated to HPF, but need additional support for message passing or task parallelism. The resulting approach can aid in making decisions on the migration path to follow, depending on the customers priorities regarding maintainability, portability, and speedup.
The methodology has been tested on two medium size applications: Simmix, a molecular dynamics simulation program and Emsim, an electromagnetic simulation program. Results of the application of the methodology and speedup characteristics of the applications will be given on a Cray T3E and on the DAS (Distributed ASCI supercomputer).