TOEP: Threshold Oriented Energy Prediction Mechanism for MPI-OpenMP Hybrid Applications

Evaluating the execution time and energy consumption of parallel programs is a primary research topic for many HPC environments. Whereas much work has been done to evaluate the non-functional behavior for single parallel programming models such as MPI or OpenMP, little work exists for hybrid programming models such as MPI/OpenMP. This paper proposes the Threshold Oriented Energy Prediction (TOEP) approach which uses the Random Forest Modeling (RFM) to train models for execution time and energy consumption of hybrid MPI/OpenMP programs. Training data (performance measurements) are reduced by ignoring code regions that have little impact on the overall energy consumption and runtime of a program and also based on the variable importance parameter of RFM. A selection parameter is introduced that selects a trade-off solution between the number of modeling points (measurement or training data) required and prediction accuracy. An exploratory study on the proposed prediction approach was employed for a few candidate hybrid applications namely HOMB, CoMD, and AMG2006-Laplace. The experimental results manifested the energy prediction accuracy of over 86.17% for large performance datasets of the candidate applications at a reduced computational effort of less than 17 seconds.