A Parallel Streamline Simulator for Integrated Reservoir Modeling on a Desktop

Abstract

Summary Desktop computing is undergoing a revolution with parallel processing on workstations. Parallel streamline simulators have been developed for shared memory architecture systems. In this paper, we discuss the implementation and performance analysis of a shared memory architecture commercial parallel streamline simulator based on native threading technology for both Windows® and Linux® operating systems. In general the streamline simulation algorithm is relatively straightforward to parallelize, however, there are several challenges that require special attention in order to avoid computing bottlenecks and inconsistent results. Repeatability of parallel simulation results is a well-known challenge. A data-accumulation scheduling algorithm designed to ensure repeatable results independent of the number of processing units has been implemented. The algorithm is supplemented by an efficient loadbalancing algorithm, to minimize processor idle time. Parallel scalability for various model characteristics and streamline solver options is analyzed. We have observed almost linear scalability up to 10 threads and a speed up factor of three to seven for simulation runs using a 1D explicit upwind finite difference solver to solve the transport problem along streamlines. The front tracking solver is inherently fast and does not gain significantly from parallelization. The run time of a serial simulation using the front-tracking solver is less than that of a parallel explicit solver run using up to 12 threads. The memory consumption of the front tracker is slightly higher than the explicit solver but overall streamline simulations are very memory efficient, enabling large geo-scale models to be simulated efficiently on 64-bit workstations. Complex three-phase blackoil simulation problems require a larger number of timesteps steps and speedup degrades as run time becomes dependent on the efficiency of the pressure solver. The results produced by our parallel streamline simulator are repeatable for any number of threads, and the algorithm used to achieve repeatability has little impact on performance.