A Framework for Multiplatform HPC Applications

This paper proposes a framework for building multi-platform applications in Java for High Performance Computing (HPC). It allows HPC developers to write their programs in Java but dynamically translate part of the programs into C programs using MPI or CUDA so that the translated code can be executed on multi-platforms. The source of the translated code is written in Java but with extensions for MPI and CUDA supports. The implementations for different platforms are switched by object-oriented mechanisms such as dynamic method dispatch. However, object oriented mechanisms are major sources of execution overheads. To reduce these overheads, the proposed framework requires that the translated code is subject to our coding rules, in which object-oriented mechanisms are available only in limited contexts. All objects except arrays must be immutable and most class types must be leaf classes. Only the types of method parameters and instance fields can be non-leaf class types. These restrictions allow our framework to statically determine object types during the code translation while they still enable building a practical class library for HPC with respect to customizability. This paper presents examples of the class libraries built on top of our framework. Their performance is sometime better than the performance of the programs written in C++ with equivalent class libraries since C++ is a general-purpose language and thus its expressiveness does not perfectly fit our problem domain, HPC applications.