Distributed Scientific Computation With JavaSpaces

Java-based tuplespa es provide a new avenue of exploration for distributed omputing with ommodity te hnology. This paper presents early results from our investigation of JavaSpa es for s ienti omputation. We dis uss weaknesses as revealed by our attempts to map existing parallel algorithms to the JavaSpa es model, and use low-level metri s to argue that several lasses of problems are not eÆ iently solvable in JavaSpa es, notably some that are demonstrably eÆ ient with other tuplespa e implementations. Nonetheless we argue that by addressing, o -the-shelf, numerous issues with distributed systems, JavaSpa es remains enti ing and may merit use for high performan e omputation if viewed less stri tly in the heritage of Linda and more as a platform-neutral ode delivery me hanism for SPMD omputing. To support this perspe tive we demonstrate parametri algorithms for whi h a JavaSpa es solution provides good speedups, relative not only to Java but also sequential C, and outline a dynami ompilation te hnique whi h for short, ompute-intensive odes already implemented in C may be used to a hieve native binary speed without ompromising the portability of the Java platform or extensively re oding the algorithm in Java. Dis ussion and empiri al results are presented in the ontext of our publi ben hmark suite, whi h we believe is among the rst available of its kind.