A CORBA-based architecture for parallel applications: experimentations with the WZ matrix factorization

This paper presents a multi-thread CORBA-based architecture for parallel applications. This architecture is hidden by a library of classes and allows the developer to make use of method invocations similar to the well known send () and receive () message passing calls. Our multi-thread architecture is built around the event nature of CORBA invocations. In order to show the pertinence and the efficiency of our architecture we present some experimental results based on an implementation of a typical parallel computation: a permuted variant of the Gaussian elimination (the WZ factorization). It appears that not only the development of parallel applications that are manageable at execution time is made easier but that in term of performances, CORBA implementation (ORBacus implementation) is comparable to the PVM one.