History Matching for Multiphase Reservoir Models on Shared Memory Supercomputers

A parallel algorithm using the generalized pulse spectrum technique and multilevel grid method for solving reservoir structural parameter identification problems is discussed here. The algorithm can be used to identify the reservoir's absolute permeability distributions by matching the computed pressure values with measured historical pressure values obtained at observation wells (history matching). Use of a multilevel grid improves the quality of the identified permeability distributions significantly. The whole parameter identification process is very computationally intensive since a group of coupled nonlinear partial differential equations (PDEs) and a regularized least square problem must be solved repeatedly with different parameter values. Several hundred megabyte memory is required for reservoir models involving thousands of grid points. The block SOR scheme with red and black ordering and a parallel Householder transformation scheme were used to solve the algebraic equations resulting from the discretization of the PDEs. High speedup has been achieved by exploring parallefisnl, refining the whole program, rather than just the hot- spots, and utilizing the high-speed cache memory efficiently.