Logo der Universität Wien

Probabilistic communication optimizations and parallelization for distributed-memory systems


In high-performance systems execution time is of crucial importance justifying advanced optimization techniques. Traditionally, optimization is based on static program analysis. The quality of program optimizations, however, can be substantially improved by utilizing runtime information. Probabilistic data-flow frameworks compute the probability with what data-flow facts may hold at some program point based on representative profile runs. Advanced optimizations can use this information in order to produce highly efficient code. In this paper we introduce a novel optimization technique in the context of High Performance Fortran (HPF) that is based on probabilistic data-flow information. We consider statically undefined attributes which play an important role for parallelization and compute for those attributes the probabilities to hold some specific value during runtime. For the most probable attribute values highly-optimized, specialized code is generated. In this way significantly better performance results can be achieved. The implementation of our optimization is done in the context of VFC, a source-to-source parallelizing compiler for HPF/F90.

Grafik Top
Grafik Top
Technical Report (Technical Report)
Scientific Computing
Institute for Software Science, University of Vienna
November 2000
Official URL
Grafik Top
Contact us
Faculty of Computer Science
University of Vienna

Währinger Straße 29
A-1090 Vienna