High Performance Scalable Computing Performance Modeling Using Ptolemy

Many of today’s signal processing applications are becoming more and more computationally intensive, requiring an increasing number of processors to satisfy their needs. Along with the larger number of processors, the bandwidth needed to pass data among the processors has also increased dramatically. In many classes of applications, designing an architecture to satisfy the data passing requirements while maintaining cost and complexity at reasonable levels is very challenging. An answer to these classes of problems is the High Performance Scalable Computing (HPSC) architecture. It consists of clusters of processing nodes interconnected over a high performance network, implemented using the MYRINET[1] protocol. Application algorithms are partitioned and mapped onto the various processing nodes in the architecture for distributed execution. The goal of this simulation work is to provide a performance modeling capability to automate the evaluation of the performance of various architectural candidates with respect to network design choices. This capability allows the development team to quickly assess the impact of various design decisions. The results of the performance simulation help in tailoring an architecture and identifying a network topology and routing scheme that best satisfies the application’s system requirements.