Evaluation and Refinement of a Tuning Tool for Grid Applications

Tuning the performance of grid applications is cumbersome because it is very difficult to decide which code to look at. In our previous work we proposed a tool to recommend tasks in a task graph that look most promising for improvement. In the present work we first show how to extract task graphs from real grid applications with the help of SCALASCA, then we evaluate the tool with the OptSched benchmark suite of synthetic schedules. Finally, we present some extensions to the tool to adapt it to some typical situations occuring in practice.