Hybrid gridding techniques provide simulation engineers with an efficient technique for refining a simulation grid in regions of the reservoir where fine detail is required. Local grids are usually required near wells to resolve coning effects. This paper describes the implementation of a hybrid gridding system where the individual local grids are processed on separate processors.
The implementation is suitable for a range of distributed memory parallel computers including a cluster of workstations linked by an Ethernet type local area network.
The basic algorithm solves the global grid with large time steps. This establishes boundary conditions for the local grids which are solved independently with small time steps. The global grid is computed on a master processor with a number of slave processors for the local grids. The algorithm allows the dynamic load balancing of the parallel work and provides a fault tolerant system for the local grids.
The system is applied to a full field simulation study. employing a network of workstations. The performance is assessed on both a dedicated cluster of machines and in a typical oil company working environment.
To understand the performance of a reservoir typically requires understanding of physical phenomena on a wide range of scales. For example, an understanding of the field pressure performance requires a full field material balance where coning effects are local to an individual well.
When the field is simulated, a range of both space and time scales often needs to be incorporated in the simulation model. One method of achieving this is by the use of a hybrid gridding technique where a coarse global grid is coupled with more refined grids near regions of interest, often wells. These types of grid systems are now being extensively used in the industry.
The separate refined grids provide a useful method of decoupling the local processes from the overall simulation. When there are many locally refined regions, for example if every well in a field is simulated with a refined grid, then the number of grid blocks in the whole simulation can be become very large. The fact that each local grid system models separate local phenomena naturally leads to the possibility of using parallel processing techniques to tackle large problems.
Simulation work in modern oil companies is typically done using a number of high power workstations connected into a local area network (LAN). Many of these machines will be idle for large parts of the day and often idle overnight. The aim of this work was to utilise these, at present, unused resources to improve the turn around time for simulation problems with large numbers of local grid refinements (LGRs).