Prediction assisted runtime based energy tuning mechanism for HPC applications

Performance tuning has become a crucial step for large-scale HPC applications, including HPC based Cloud applications. A need for an energy-aware autotuning solution has recently widened research thrusts among energy conscious scalable application developers. There exist a few standalone energy reduction approaches such as reducing MPI wait times, diligently selecting CPU frequencies, efficiently mapping workloads to CPUs, and so forth for HPC applications. Implementing energy-aware autotuning mechanisms for HPC applications, however, might require multiple executions if exhaustively tested. This paper proposes a prediction assisted energy tuning mechanism named Random Forest Modeling based Compiler Optimization Switch Selection mechanism (RFM-COSS) for HPC applications. RFM-COSS was implemented using RFM algorithm and its variants, namely RFM-SRC and RFM-Ranger. The training datasets of RFM-COSS were created using DiscretePSO algorithm for a few candidate benchmarks such as hpcc, MPI-Matrix, and Jacobi. The experimental results of the proposed RFM-COSS prediction mechanism achieved 17.7 to 88.39 percentage points of energy efficiencies for HPC applications.