On December 15th 2017, Dr. Robert Mijakovic passed his dissertation defense with his thesis titled “Automatic Online Tuning of HPC Applications”. Congratulations!
The abstract of the work is:
Performance tuning of scientific codes often requires tuning many different aspects like vectorization, OpenMP synchronization, MPI communication, load balancing. The Periscope Tuning Framework (PTF), an online automatic tuning framework, relies on a flexible plugin mechanism providing tuning plugins for different tuning aspects. To reach the exascale level within given energy constraints, system vendors rely on a wide range of more energy-efficient accelerators and manycores. This work focuses on enabling tuning plugins for such state-of-the-art HPC architectures and on their combination by automatically selecting plugins based on the prediction of their tuning potential. Two special tuning plugins were designed and implemented in this thesis that support tuning of OpenCL codes based on selecting the best combination of flags for offline compilation of kernels as well as on the investigation of application level tuning parameters specified by the user. The thesis also presents the concept of meta-plugins that combine individual tuning plugins. Since each plugin can take considerable execution time for testing the various combination of the tuning parameters, it is desirable to automatically predict the tuning potential of plugins for programs before their application. We developed a generic automatic prediction mechanism based on machine learning for this purpose. This thesis demonstrates this technique in the context of the Compiler Flags Selection and the MPI Parameters plugin, which tune the flags of the compiler and the parameters of the MPI library, respectively. When the tuning process finishes, each tuning plugin generates a tuning advice that is automatically applied either before and during the runtime of the application.