Applications experience in Jade

This paper presents our experience developing applications in Jade, a portable, implicitly parallel programming language designed for exploiting task-level concurrency. Jade programmers start with a program written in a standard serial, imperative language, then use Jade constructs to describe how parts of the program access data. The Jade implementation analyzes this information to automatically extract the concurrency and execute the program in parallel. The resulting parallel execution is guaranteed to preserve the semantics of the serial program. We have implemented Jade as an extension to C on shared memory multiprocessors, a homogeneous message-passing machine and networks of heterogeneous workstations. To evaluate Jade, we obtained several complete scientiic and engineering applications and parallelized them using Jade. We then executed these applications on several computational platforms. We use this applications experience to evaluate Jade with respect to two properties: how well Jade supports the process of writing parallel programs and how well the resulting programs perform. Our applications experience shows that the current version of Jade is a qualiied success. For all but one application the use of Jade entails limited programming overhead. The coarse-grain computations perform very well, with the dynamic Jade overhead having no signiicant impact on the performance. The ner-grain computations suuer from some Jade-speciic performance problems, but some of these could be eliminated with a more advanced Jade implementation.