Towards Efficient Compilation of the Hpjava Language for High Performance Computing

This dissertation is concerned with efficient compilation of our Java-based, highperformance, library-oriented, SPMD style, data parallel programming language: HPJava. It starts with some historical review of data-parallel languages such as High Performance Fortran (HPF), message-passing frameworks such as p4, PARMACS, and PVM, as well as the MPI standard, and high-level libraries for multiarrays such as PARTI, the Global Array (GA) Toolkit, and Adlib. Next, we will introduce our own programming model, which is a flexible hybrid of HPF-like data-parallel language features and the popular, library-oriented, SPMD style. We refer to this model as the HPspmd programming model. We will overview the motivation, the language extensions, and an implementation of our HPspmd programming language model, called HPJava. HPJava extends the Java language with some additional syntax and pre-defined classes for handling multiarrays, and a few new control constructs. We discuss the compilation system, including HPspmd classes, type-analysis, pre-translation, and basic translation scheme. In order to improve the performance of the HPJava system, we discuss optimization strategies we will apply such as Partial Redundancy Elimination, Strength Reduction, Dead Code Elimination, and Loop Unrolling. We experiment with and benchmark large scientific and engineering HPJava programs on Linux machine, shared memory machine, and distributed memory machine to prove our compilation and proposed optimization schemes are appropriate for the HPJava system. Finally, we will compare our HPspmd programming model with modern related languages including Co-Array Fortran, ZPL, JavaParty, Timber, and Titanium.