Applications based on highly irregular, dynamically changing data structures cannot be handled efficiently with current High Performance Fortran and available HPF compilers. In order to support such applications, extensions of the language as well as sophisticated analysis and parallelization techniques are required. In this paper we introduce the Vienna HPF+ Compiler (VFC), a new source-to-source parallelization system for HPF+, an optimized version of HPF, which addresses the requirements of advanced irregular applications. In addition to extended data distribution and work distribution mechanisms, VFC supports various features of HPF+ to reduce preprocessing and communication overheads. The system provides powerful runtime parallelization strategies for complex loop nests with indirect array accesses which are common in irregular codes. Performance measurements show that the use of language features for communication schedule reuse results in a performance close to hand-written message passing programs.

Top