Henrik Löf
Division of Scientific Computing
Department of Information Technology
Uppsala University
Fluid flow simulators for oil reservoirs are being used today not only to optimize oil and gas production but also to investigate the feasability of carbon sequestration. The equations that govern such sub-surface flows are often very challenging to solve because of the complex interactions between geological properties and multi-phase fluids . The size of the simulation domains and the multi-scale character of the non-linear models also lead to computationally very costly solution procedures which can be addressed using techniques like adaptive mesh refinement and parallelization.
In heterogenous reservoirs the time scale at which fluids flow along streamlines is generally much shorter than the scale at which the streamlines change. This observation motivates the decoupling of the transport problem into a sum of one-dimensional problems along streamlines. In this talk I will discuss algorithms for solving these Euler-Lagranian-type methods efficiently in parallel using a shared memory programming model. Although the 1D equations are completely uncoupled they couple to the pressure equation used to generate the velocity field in which the streamlines live. Furthermore, in contrast to many other numerical solution methods, streamlines are not grid-local operations and they present a dynamic, and variable workload because of well placements and heterogenous rock properties. We try to address these challenges using data locality optimizations and load balancing algorithms to produce efficient streamline simulators for modern multi-core systems.
This work is in collaboration with Margot Gerritsen, Stanford University and Marco Thiele, Streamsim Technologies, Inc.