A Non-Linear GPU Performance Modeling Approach and Consolidated Linear Hardware Model Performance Evaluation of the LEAP Cluster

This document discusses the evaluation of a consolidated linear performance model of Texas State University's HPC LEAP Cluster, and the development of a non-linear model to represent the performance of the NVIDIA V100 Graphics Processing Unit (GPU). The consolidated model constructed for the LEAP Cluster was evaluated by using it to compute execution times for different sizes of processes and comparing those times to the actual processes run on LEAP. The model itself is found to effectively represent the performance trends of LEAP but fail to match the actual value. The GPU's performance analysis was done by first collecting performance data then using non-linear methods to model it. The number of floating-point operations required to perform a square matrix multiplication is readily known, so the performance data was gathered by running a simple square matrix multiplication program and measuring its execution time. The size of the matrices was then used to determine the number of floating-point operations and divided by the execution time to calculate FLOPS. The final model for the GPU was generated by using a set of logarithmic functions that mapped the GPU performance to a semi-logarithmic scale of the number of floating-point operations.