Parallel Streamline Simulation
- Henrik Tobias Löf (Stanford University) | Margot Geertrui Gerritsen (Stanford University) | Marco Roberto Thiele (Streamsim Technologies, Inc.)
- Document ID
- Society of Petroleum Engineers
- Europec/EAGE Conference and Exhibition, 9-12 June 2008, Rome, Italy
- Publication Date
- Document Type
- Conference Paper
- 2008. Society of Petroleum Engineers
- 5.5.8 History Matching, 5.5 Reservoir Simulation, 5.5.7 Streamline Simulation, 5.4.2 Gas Injection Methods, 5.4.1 Waterflooding, 1.6.9 Coring, Fishing, 5.3.2 Multiphase Flow, 4.3.4 Scale
- 4 in the last 30 days
- 684 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 5.00|
|SPE Non-Member Price:||USD 28.00|
We present an efficient strategy for parallelizing streamline (SL) simulators using shared-memory programming models and evaluate its performance. This is the first work to extensively address parallelization of streamline solvers. We consider sharedmemory programming models because of their suitability for SL simulation and because of the need to sustain peak performance on the multi-core processor platforms of the future. We test our implementation on a 8-way Sun Opteron server, representative of the
state-of-the-art shared memory systems in use in the industry, as well as the very recently released Sun Niagara II multi-core machine that has 8 floating point compute units on the chip.
We consider both static and dynamic load-balancing algorithms and take into account associated communication overhead and data locality issues. We apply our experimental parallel SL simulator to two examples: the first test case is taken from the 10th SPE comparative solutions project, and the second is representative of large industrial reservoir applications. We use simple, single-phase, linear flow physics in order to clearly distinguish between system-related and physics-related run time performance.
We found that by using a large amount of streamlines per thread the dynamic load balancing strategies built-in to OpenMP performed well as the average workload per thread is not very sensitive to workload variations between individual streamlines. On the Opteron system, we find parallel efficiencies ranging between 60 and 75 %. The sublinear speedup is mostly due to a low degree of data locality in the tracing stage and communication overhead in the mapping stage. In applications with more complex physics, the
relative contributions of these stages will decrease. Thus we expect the parallel performance, which is already attractive, to be higher in most industrial applications. On the Niagara II, we obtain almost perfect linear scalability thanks to the lowered communication costs on these architectures. This result is all the more satisfactory considering that future server designs will be akin to this system.
A frequently mentioned computational advantage of streamline methods is that the decoupled one-dimensional transport systems defined along each streamline are naturally parallelizable. For many displacement processes where streamlines have proven to be particularly effective, such as water flooding (Thiele and Batycky 2006), compositional modeling (Gerritsen et al. 2007), and reservoirs described by dual porosity models (Di Donato et al. 2003) the bulk of the computational work is associated with solving the relevant transport equations along the streamlines. Although the assumed inherent, high-level parallelism of streamline simulation is often given as one of its appealing feature, to date little work has been published on this.. The principle motivation for our work is therefore to discuss parallelization of streamline simulation and study the challenges facing such an implementation. While it is true that the decoupled mass balance equations that are solved along the streamlines are independent of each other and therefore naturally parallelizable, good parallel performance requires careful attention to load balancing, computational overhead, memory locality issues and costs associated with communication. We address these issues in our work.
Our second motivation for studying parallel streamline simulation is the requirement by modern reservoir engineering workflows for an ever-increasing number of simulations. With a parallel streamline implementation we envision significant advances in optimization, history matching, uncertainty, and classic sensitivity analysis. Being a fast proxy, streamline simulation has already made a significant contribution to a more efficient and useful application of reservoir simulation to reservoir management. A parallel
implementation of streamline simulation should further increase the added value that streamline simulation already brings to reservoir simulation, engineering and management.
|File Size||423 KB||Number of Pages||8|