Java Grande and HPJava

Investigators: B. Carpenter, G.Fox

Following a workshop organized by Fox in December 1996, there has been growing interest in the use of Java and its associated technologies in large-scale computation. The workshop series is popular and now well established with ACM sponsorship. Further, the Java Grande Forum (http://www.javagrande.org) established in March 1998, has divided the issues into two broad areas: Numerics and Concurrency. The Numeric subgroup led by the NIST group of Boisvert and Pozo has brought to the Java standards process a set of proposals which are designed to make the Java language appropriate for core science codes. Clearly Java can be used client side for nifty animations and the use for modern servers is well established in both industry and academia. However can Java be effective in your basic parallel algorithm? Satisfying this goal brings up issues such as floating point performance, convenient expression of complex variables and efficient array constructs. Already the Java Grande Forum, coordinated by Fox, has had significant impact on Sun Microsystems. As well as these community activities, we focus in CSIT on several specific research projects. Later we discuss work on web-based Problem Solving Environments and Science portals while here we describe research on new approaches to Java-based data parallel programming. This is a collaboration with Professor Xiaoming Li from Peking University with whom we developed the basic ideas as part of research on compiler runtime for High Performance Fortran and C++. The HPJava model is less ambitious than systems like High Performance Fortran (HPF) and aims to support an SPMD model intermediate between basic message passing (MPI) and HPF. We call the generalization to languages like Fortran and C++ HPspmd but at Florida State we focus on Java. HPJava can be implemented as either threads or distributed MPI and we would be early users of both the proposed major parallel architectures (in the research infrastructure) to compare the effectiveness of these two paradigms. Currently HPJava has a few key concepts: distributed arrays, data-parallel arrays and special control constructs allowing one to restrict either the processors or array elements used. There is an elegant syntax for user specified decomposition. Currently we are testing the first HPJava compiler (really just a translator) but we will then take some ab-initio application codes (such as those in our GEM Earthquake activity) and see if the new language can produce good code in a practical case. Obviously this can drive changes to both the implementation and the basic parallel concepts.

As part of the pure MPI syntax for HPJava, we have developed an MPI Java binding which is available for download (http://www.npac.syr.edu/projects/pcrc/HPJava) and is a reference implementation used by the Java Grande MPI subgroup. We are developing an interesting research thrust looking at MPI style messaging which is natural for languages like Java. This involves rather obviously serialized object based messaging replacing derived types in MPI. However more interesting is links to technology like Jini to support dynamic invocation of parallel codes on cluster architectures. We hope this will be a major thrust on the proposed cluster resource. Here we link to work described later on Science Portals which are the natural software model for the user interface and middleware linking parallel Java codes to data and other services.

We hope that the next few years will see a convergence of the standards work (improved numerical performance of Java) and research like ours (parallel programming paradigms exploiting features of languages like Java). The net result should be more attractive programming environments for computational science.