A Tool for Performance Analysis and Prediction of Parallel Computing on NOW

摘要

In this paper, we study performance measurements and predictions for parallel computing on a homogeneous network of workstations (NOW). The knowledge of program execution times is crucial for the development and the verification of software. The PMPtool's approach is based on a two-level model, where on the top a semi-deterministic task graph is used to capture the parallel execution behavior of a given application, and on the bottom level, it is extracted system and parallel program characteristics through its execution. Data are input to a prediction tool, where we can predict results from another system with different system characteristics, such as network latency and badwidth, CPU speed. We validate the prediction model using experiments on a homogeneous network of workstations.