Optimizing Performance of Automatic Training Phase for Application Performance Prediction in the Grid

Automatic execution time prediction of the Grid applications plays a critical role in making the pervasive Grid more reliable and predictable. However, automatic execution time prediction has not been addressed due to the diversity of the Grid applications, usability of an application in multiple contexts, dynamic nature of the Grid, and concerns about result accuracy and time expensive experimental training. We introduce an optimized, low-cost, and efficient yet automatic training phase for automatic execution time prediction of Grid applications. Our approach is supported by intraand inter-platform performance sharing and translation mechanisms. We are able to reduce the total number of experiments from an polynomial complexity to a linear complexity.