A Performance Prediction Methodology for Data-dependent Parallel Applications

The increase in the use of parallel distributed architectures in order to solve large-scale scientific problems has generated the need for performance prediction for both deterministic applications and non-deterministic applications. In particular, the performance prediction of data-dependent programs is an extremely challenging problem because for a specific issue the input datasets may cause different execution times. Generally, a parallel application is characterized as a collection of tasks and their interrelations. If the application is time-critical it is not enough to work with only one value per task, and consequently knowledge of the distribution of task execution times is crucial. The development of a new prediction methodology to estimate the performance of data-dependent parallel applications is the primary target of this study. This approach makes it possible to evaluate the parallel performance of an application without the need of implementation. A real data-dependent arterial structure detection application model is used to apply the methodology proposed. The predicted times obtained using the new methodology for genuine datasets are compared with predicted times that arise from using only one execution value per task. Finally, the experimental study shows that the new methodology generates more precise predictions