Performance modeling of adaptive computing systems

Adaptive Computing Systems (ACS) are computer based systems that can change their architecture in response to changing application requirements. In this research, the system is made adaptive by designing multiple modes of operation, each represented by a dataflow model of computation at the algorithm level and providing a run-time support environment that enables dynamic switching among the different modes in response to changes in the environment or changes in the computational requirements.; A performance modeling framework is essential in understanding and designing the ACS system. The framework developed in this research enables rapid performance evaluation of multiple design alternatives, and operating modes, by separating functional behavior from performance behavior. Two modeling levels are considered: An algorithm performance model that only focuses on modeling the application and a resource performance model that models the application, compute platform and the mapping from algorithm to architecture. Both models are fairly high on a pre-characterized library of performance primitives. Both hardware and software aspects are uniformly modeled using VHDL. Careful choice of the modeling style is essential to achieve fast simulations by minimizing unnecessary simulation events.; The Automatic Target Recognition (ATR) System is used as a case study to show how the performance modeling framework can be applied. The experiments clearly demonstrate the effectiveness of the modeling framework in understanding and optimizing system performance issues. These issues include finding the bottlenecks in the algorithm dataflow, mapping from algorithm to different architectures, effect of communication link bandwidth, architectural features of the target processors as well as effect of reconfiguration on overall system performance.