Abstract:

[Objective] This paper studies scalable black-oil simulation on domestic deep computing units (DCUs) for complex real corner-point models with about one thousand wells and tens to hundreds of millions of cells, focusing on partitioning that preserves flow topology and near-well coupling for multi-well operations, as well as heterogeneous fully implicit solution on domestic DCUs. [Methods] A physical-connectivity graph containing Cartesian-face, pinchout/non-neighboring-connection (NNC), fault-juxtaposition and well-cohesion edges is constructed to generate well-affinity partitions. Around the constrained-pressure-residual (CPR) preconditioning framework, fine-grid sparse operations, halo exchange, reservoir-side assembly, convergence reduction, and matrix filling are organized for DCU execution. [Results] On a ten-million-cell, thousand-well field case partitioned into 64 subdomains, well-affinity partitioning reduces Allreduce synchronization in well-event steps by about 34%-43% relative to structured partitioning, and the cohesion-edge variant gives the lowest cumulative Allreduce and MPI wait time. On a hundred-million-cell case partitioned into 400 subdomains, under a fixed 400-way decomposition and identical MPI layout, DCU-side execution achieves about 6.6×full-run speedup over a single-threaded CPU-rank path; steady-step execution reaches about 5.66×speedup and water-injection event steps reach about 3-10×speedup. [Conclusions] Well-affinity partitioning based on real flow topology and perforation coupling reduces global synchronization in well-event-dominated operation while increasing bounding-box and point-to-point communication costs. The DCU-oriented heterogeneous solve shows stable gains across complete report steps, while cell-wise numerical cross-checks, strongly nonlinear event steps, first-step warm-up cost, and complete full-schedule runs at larger scales require further validation.

Key words: black-oil model, reservoir simulation, well affinity, domestic DCU, heterogeneous computing