Research On Efficient Parallel Computing For Ground Water Flow Simulation

The research on parallel numerical groundwater simulation is of great practical significance in solving the environment hydrogeological and engineering geological problems. As researches go further, urgent requirements have been posed on modeling groundwater with much-refined spatial and temporal grids. However, the traditional serial methods have reached their limits in solving large-scale problems due to the huge computational ability requirements, large memory capacity requirements and the restriction on the data channel bandwidth. This phenomenon has greatly hindered the application of groundwater numerical simulation; and thus it is urgent to develop an efficient method for large-scale groundwater modeling.This dissertation mainly devotes to developing an efficient parallel computing method for groundwater flow simulation. It focuses on using some advanced theories on the parallel solution to the partial differential equation, combined with some up-to-date achievements in CPU cluster, multi-core shared memory and the GPU-CPU system, to overcome the performance limitations in solving the linear system arising from the discretization of groundwater flow equations. The main contributions of this dissertation are summarized as follows:(1) The parallelism in the groundwater flow modeling process has been extensively analyzed. Based on different parallel architecture, reasonable mapping strategies between the dominant computing tasks and different levels of corresponding architectures are presented. This has provided a theoretical proof for designing efficient algorithm structures, support structures and implementation details. The experiment results demonstrate that, the coarse-grained data parallel strategy adopted on distributed memory system and the fine-grained multi-threading method presented on shared memory as well as the GPU-CPU system, have achieved desirable performance.(2) The optimal selection and construction of Krylov subspace method, preconditioner as well as their combination are provided for groundwater flow simulation. On distributed memory system, we present a combination of different preconditioners for better performance, which is the blocked iterative algorithm for the whole partitioned domain and incomplete factorization method for each subdomains. The experiments reported in the paper confirmed that the combination of block Jacobi and ILU(0) is the best performing preconditioner and has good scalability. On GPU-CPU system, the combination of Jacobi preconditioner and CG algorithm is adopted to solve the linear system. The experiment results support our claim that this combination has a good trade-off between the cost of constructing and applying the preconditioner, and its gain in convergence rate.(3) A novel assembly algorithm for sparse matrix arising from the discretization of groundwater flow equations has been presented and an interface for linking MODFLOW and general parallel iterative solvers has been developed. CSR format is used for data transmission to reduce the communication overheads between CPU and GPU; BCRS block algorithm is presented for better utilization of GPU registers and further to optimize the sparse matrix- vector multiplication operation.Other important works include: a parallel infrastructure with hierarchy design and good scalability has been proposed for groundwater modeling; a pre-processing and post-processing method for hydro-geological data has been given in solving the large-scale three-dimensional groundwater model.