Section: New Results

Performance model for multithreaded applications on multi-core processors

Concerning data locality, researches have shown a tradeof in groupement strategy for process mapping. We have to deal with balanced improvement of several aspects such as threads synchronizations or resource exploitation. Weighting those criterias can only be achieved according to a certain knowledge of both the application and the machine.

Thus, we are working on modeling threads affinity and weights on machines topology to improve a placement method based on the TreeMatch algorithm using new metrics. Several experiences have lead us to the conclusion that it is very hard to identify the key hints and to understand application needs.

Consequently, we are developping a visual tool which displays hardware counters aggregated and mapped on the system topology to identify dynamically those hardware narrows during execution, and understand processes placement effects on them. We hope to achieve a better comprehension of process placement consequences on resources usage by applications.