A methodology for performance prediction of Sphinx I in multi-computer architectures

A methodology is proposed for performance prediction of an application in a multi-computer architecture. The methodology introduces a two-step approach. In the first step the application (Sphinx I) is considered as a benchmark and its performance in uniprocessor systems is predicted from other standard benchmarks. In the second step the predicted performance values for the compositive modules of the application are used in a partition model to predict the performance of the entire application in a multi-computer architecture. The approach is validated by measuring the actual performance of the application on different two-processor configurations and comparing the results to the predictions made. The methodology for predicting application performance in a multi-computer architecture provides a relatively inexpensive approach to design. Rather than porting the application to the target architecture to measure its performance, the performance can be estimated from standard benchmarks, such as Dhrystone and Linpack. Results from Dhrystone and Linpack benchmarks are readily available, even for new systems, since the results are typically published in advertising literature.<
>