A Performance Prediction Model for Structured Grid Based Applications in HPC Environments

Predicting the performance of parallel applications at scale is a challenging problem. We have developed a performance prediction model for structured grid-based scientific applications for High Performance Computing systems. Our model can capture the system complexity and consider computation and communication attributes of application performance on HPC architectures. We have also proposed a methodology for obtaining the realistic value of the model parameters by small-scale sample runs of an application on the target system. We have used our model to predict the performance of an actual application (2D Flood Simulation) and a synthetic application (Game of Life) on the Summit supercomputer at Oak Ridge National Laboratory and Stampede2 at Texas Advanced Computing Center. The experimental results indicate that our model’s predictive performance is acceptable, and our model was able to predict the performance of these two applications on Summit and Stampede2 with more than 90% accuracy.