| Abstract |
|---|
| High Performance Fortran (HPF) is the de facto standard languagefor writing data parallel programs. For distributed memoryarchitectures, HPF compilers emulate the global address spaceby distributing the arrays among the processorsaccording to the mapping directives of the userand by generating automatically explicit interprocessor communication.In case of applications that use indirect addressingon distributed arrays, HPF compilers have limitedcapabilities for optimizing such codes,especially for optimizing communication andreusing communication schedules between subroutine boundaries.This paper describes a dynamic approach for optimizing unstructuredcommunication in codes with indirect addressing. The basic idea is thatruntime data reflecting the communication patterns will be reusedif possible. The user has only to specify which data in theprogram has to be traced for modifications. The experiments and resultsshow the effectiveness of the chosen approach. |
| Contact |
| Thomas Brandes Institute for Algorithms and Scientific Computing (SCAI),German National Research Center for Information Technology (GMD),P.O. Box 1319,D-53754 St. Augustin, West Germany Thomas.Brandes@gmd.de |