Parallelisation of an oil reservoir simulation

Abstract

The importance of oil reservoir simulation is self evident. Almost every device with a moving part uses oil for lubrication, most forms of transport rely on oil as a source of power and many manufacturing processes use oil as a raw material. Stocks of oil are limited and our ability to recover the available oil is dependent on our knowledge and understanding of the oil reservoir. Simulating oil reservoirs is a problem which requires large amounts of memory and computation. Models typically consist of 50 to 100 thousand grid blocks each of which could typically require 2KB of data storage during the simulation. Typically simulations may take hours or even days to run. To achieve even this each cell may have areal dimensions in the region of 100m much larger than the typical length scales in the reservoir so that rock properties have to be up-scaled from those measured using seismic and core data. Clearly, then there are good reasons why it is desirable to be able to run significantly larger models or existing models on a shorter time scale. Eclipse 100 is arguably the most widely used oil reservoir simulator. A “scalable” parallel version of the code was developed under the auspices of the EUROPORT2 project. The key to the parallelisation of this oil reservoir simulator was the development of a new preconditioning algorithm for the conjugate gradient method which forms the heart of the simulator. In this paper we describe briefly how the new preconditioning algorithm was developed from the original serial preconditioner and present some performance figures for 3 data sets.