Research and Application on GPU Parallelization of Reservoir Simulation Software

Improving the computing speed is one of the technical development directions of reservoir simulation software. The characteristic of GPU is to process intensive data and parallel data computation, It has obvious advantages over CPU in floating-point computing processing capacity and memory bandwidth, and has certain advantages in cost and power consumption. Now the performance of a single high-end GPU has reached the performance of the traditional multi-core CPU cluster. CUDA technology based on GPU makes it possible to significantly improve the speed of reservoir simulation. DQDUAL software is a simulator for low-permeability reservoirs developed by Daqing Oilfield, and its speed and scale can not meet the simulation requirements. Based on the source code of DQDUAL simulator, this paper developed an interface program for linear solving of the simulator using the AmgX calculation library provided by NVIDIA, converts the linear sparse matrix from DIA format to CSR format, successfully migrated the simulator from CPU platform to GPU platform, and realized GPU computing solution for linear algebraic equation set of DQDUAL simulator. The main calculation modules of DQDUAL simulator are developed for GPU parallelization by using CUDA Programming, and the asynchronous transmission is optimized to achieve the CPU + GPU parallel calculation of DQDUAL simulator. Through four practical model examples, the new version of DQDUAL simulator has all correct results, with linear solving speed increased by more than 10 times, and the speed of the simulator increased by more than 6 times, and the speed of some models can be increased by 30 times. This project confirms that existing reservoir simulation software can be upgraded to GPU parallel version through code transformation, but the speed improvement after that is not upgrading according to GPU design framework is lower than expected. A large number of CPU-based industrial reservoir simulation software have the demand for speed improvement through technological breakthroughs. This study provides valuable reference for GPU parallelization transformation of existing simulation software.