Future-safe performance
There is a significant computational and memory performance gap between GPU-type hardware and CPUs, and most new HPC systems have GPUs as their main building block. To remain relevant as a high-performance solution, any computational software must not just be parallel, but also able to be deployed on GPUs with good efficiency.
Method and approach
The linear solver part is the most difficult subsystem in the simulator to port to the GPU with high performance. Building on experience and implementations from, e.g., the ACROSS project, we extend and improve the modular and efficient gpu-ISTL framework for linear solvers to provide a high-performance implementation of a full CPR+AMG linear solver stack on the GPU. The aim is also to port the remaining most performance-critical parts of the simulator (i.e. property evaluation and linearization) to the GPU while minimizing separate code paths for CPU and GPU.
OPM Flow is free software
All products of the project such as software improvements or documentation are shared freely with the community using free and open licenses. The work is done following an open and collaborative development model, where any user or developer of OPM Flow can discuss changes and improvements, as well as testing and using them without restriction. This enables simple cooperation between industrial and research communities for mutual benefit. Extensive testing on industrial assets yields robustness and performance benefits for all users, not just the asset owner.