SC Conference - Activity Details

A Performance Prediction Framework for Multi-core Applications

Ana Lucia Varbanescu  (Technical University Delft)
Doctoral Research Showcase Session
Thursday,  11:00AM - 11:15AM
Room 17A/17B
The selection of an appropriate multi-core platform is difficult for high performance computing applications: peak performance requires many architecture-specific and non-portable application code transformations. Therefore, tuning the application for every multi-core platform in order to find "the right one" is impractical. My thesis focuses on finding an efficient strategy for matching arbitrary applications to multi-core platforms. I propose a performance prediction framework based on three elements: a parameterized application task-graph, an architecture model, and a mapping between them. The models are used by an analytical engine to give a performance estimate and/or feedback information for tuning the models. The experimental results show that the framework can be used to correctly match a large variety of applications to a representative set of multi-cores. To further reduce the modeling efforts, I am currently generalizing this strategy towards matching classes of applications and platforms.
   IEEE Computer Society  /  ACM     2 0   Y E A R S   -   U N L E A S H I N G   T H E   P O W E R   O F   H P C