Data Parallel SPMD Programming Models from Fortran to Java

点击翻译按钮获取中文摘要

This project brings together several important research areas, includingparallel compilers, Single-Program Multiple-Data (SPMD) libraries, andobject-oriented programming models. The project combines these ideas toachieve high performance on parallel machines in a robust fashion over arange of languages.Building on earlier research in the Parallel Compiler Runtime Consortium(PCRC) project, this project will investigate a language model thatcombines data-parallel features from High Performance Fortran (HPF) with anexplicitly parallel SPMD programming style. This "HPspmd" model allowsdirect calls to many established libraries for parallel programming withdistributed data. By relaxing HPF's single-threaded semantics and havingthe programmer explicitly manage non-local references, the task of thecompiler is hugely simplified. However, good library bindings (e.g. classlibraries) supply the required expressiveness for programming. Thisproject will develop and optimize a Java version of the language model(HPJava); design and implement bindings to parallel libraries such asCHAOS, DAGH, Global Arrays, MPI,and ScaLAPACK; and implement largeapplications within the resulting framework. The implementations areexpected to drive further refinement of the language model, especially insupport of irregular problems.