SIMPAR: A Portable, Object-Oriented, Simulation-Science based, Metamodel Framework for Performance Modeling, Prediction, and Evaluation of HPC Systems

摘要

We present a novel, portable, platform-independent, object-oriented, simulation-science-based, metamodel framework (SimPar) for performance evaluation, estimation, and prediction of High-Performance Computing (HPC) systems. This UML-based, parallel meta-model enhances the Bulk Synchronous Parallel (BSP) computation model. The UML activity diagram is used to model the computation, communication, and synchronization operations of an application. We also identify the UML building blocks that characterize the message passing and shared memory parallel paradigms. This helps in modeling large and complex parallel applications. Using the collaboration diagram concept, parallel applications are mapped onto different multiprocessor architecture topologies such as hypercube, 2D mesh, ring, tree, star, etc. We present unique UML structural and behavioral extensions for modeling the inter-object interactions in BSP model. The communication semantics such as BROADCAST, GATHER, and SCATTER are incorporated in the metamodel using UML building blocks. In its present form, UML cannot satisfy all the modeling needs. In addition, none of the currently available tool sets deploy UML-based modeling. This underscores the uniqueness of parallel, cluster-based UML-enhanced framework presented here. We have validated the proposed model through benchmarks, simulation-science case studies and real-time parallel applications.