Harnessing Data Movement Strategies to Optimize Performance-Energy Efficiency of Oil & Gas Simulations in HPC

The computing demands of Oil & Gas exploration applications have increased over the years. As a result, graphic processing units (GPUs) are becoming an essential resource due to their high-performance capabilities. At the same time, energy consumption has emerged as a significant challenge as the power requirements of such systems are increasing according to data needs. Because Oil & Gas applications have many data exchange points between CPU and GPU memories to keep data updated during the execution, improving the communication task is essential to enhance the trade-off between performance and energy consumption, represented by the energy-delay product (EDP). Hence, in this paper, we employ four different data movement optimization strategies on an RTM application to improve performance, energy, and EDP. Through extensive experiments over five different NVIDIA GPU Products (from Pascal to Hopper architectures, and NVIDIA GH200 Superchip), we show that employing the right data movement strategy can improve performance up to 62.2% and EDP up to 78.1%. We also shown that advances in the software and hardware layer of NVIDIA GPUs over generations have positively impacted the unified memory technology in terms of performance, energy, and EDP.