9411394 Berry This research proposal focuses on the development of a reasonably large suite of real scientific application programs for execution on networks of high-performance workstations or within a limited distributed computing environment with a relatively fast network (e.g., 32-processor Thinking Machines CM-5). This suite of Scientific Applications for Distributed Instrumentation and Execution of SADIE can be used a testbed of real scientific applications by computational scientists to evaluate and benchmark software tools or environments for various distributed computing environments. SADIE will facilitate the development of performance prediction models which can assess the suitability of a given environment for a particular application or subset of applications. Realistic memory, CPU speeds, and network latencies needed to execute scientific applications will be quantified using several distributed computing software systems, Candidate support software systems for SADIE include PVM, P4, CONDOR, and Linda. By constructing a testbed of applications, the effectiveness of popular (underlying) algorithms which previously performs well on massively-parallel platforms can be assessed. Alternative strategies/algorithms for distributed environments can be compared so that a classification in which algorithms are ranked (in suitability) according to certain network conditions is produced. SADIE comprises a diverse set of scientific applications: modeling high-temperature supercomputers, modeling piezoelectric crystals, first principles simulation of material properties, large-scale information retrieval, computational ecology, and finite element computations on unstructured meshes. The corresponding set of dominant algorithms used within SADIE includes: the solution of linear systems, SVD and generalized symmetric eigenproblems, adaptive mesh refinement, matrix assembly from finite element meshes, and cluster/patch identification geometry.
