Currently running on our 32-node CM-5E, POWERS is much faster than conventional simulators running on more expensive vector supercomputers, such as the Cray-2 and Cray-YMP. An early study, using a conventional 120,000 cell model of an offshore field, showed that we could simulate 26 years of production history in approximately 5 minutes. Using a more appropriate resolution of 1.12 million cells, we were able to run a 26-year history in approximately three hours. This made running high-resolution models practical for the first time. At the June 1996 SPE Technical Symposium, we presented our results to date in Simulation of Berri-Hadriya Reservoir History with POWERS.
The speed advantage of POWERS continues to grow as the number of parallel processing nodes is increased. The large model listed above takes about 1 hour on a 128 node CM-5E. Performance of POWERS on machines up to 512 - 1024 nodes is believed to be good, based on benchmarks of the linear algebra kernel. Scalability of the application under cmf & HPF is inherently good, because property data is distributed in a way that minimizes communication between nodes.
The faster processing speed of POWERS makes it practical to run models that have much higher spatial resolution than is current practice. Due to the speed and memory limitations of vector supercomputers, the huge oil reservoirs at Saudi Aramco are insufficiently resolved by using approximately 14 - 20 vertical layers, and a horizontal resolution of 1-km. (Yes, that is kilometer!) These low-resolution models require a disturbing about of "adjustment" to calibrate against historical pressure & water production.
The higher resolution models available to users of POWERS reduce the need to fudge the geological data in order to calibrate. The result is greater confidence in the predictive power of the simulation, and a dramatic reduction in cycle time. You need to trust the results of simulation AND get them quickly in order for computer modeling to be of practical use in decision support.
The problems associated with the larger data-sets required for high resolution models have been overcome. POWERS exploits parallel IO to avoid bottlenecks in writing visualization and restart data. Pressure, saturation and well rates generated by the simulator can be viewed during and after the simulation run. Typical delay times for data preparation are less than two minutes, even with NFS congestion.