Who's Fastest?

In 2005, Landmark Graphics announced that they had developed a simulator that was "five times faster than existing technology".  We have some evidence that indicates "five times slower" would be much closer to the true number.

In 2004, developers of Landmark's new simulator reported its performance¹ on the black oil industry benchmark problem from the Ninth SPE Comparative Solution Project (SPE9)².  The reported cpu time on a 3.05 GHz PC is 27.1 seconds, with 17.0 seconds in the linear solver.  Sensor’s cpu time on a 2.8 GHz PC of 2004 vintage* is 8.2 seconds, with 5.0 seconds in the linear solver.  Sensor data and output files are spe9.dat** and spe9.out.  Sensor is 3 or 4 times faster than Nexus on this problem, in both solver and non-solver time.  On a per-Newton iteration basis, Sensor is 2 or 3 times faster, but Nexus requires 32% more of them to complete the run (56 for Sensor, 74 for Nexus).  VIP was reported as being only 30% slower than Nexus and took 112 Newtons.  In terms of cpu time per Newton iteration, Nexus was shown to be no faster than VIP in either total time or solver time.

Of the eight participants in the 1995 SPE9 Comparative Solution Project, Sensor was reported as having the fastest time².  Sensor's superiority was prominently noted in the 1995 paper : "First, the formulation in the SENSOR program does allow a significant reduction in the total number of Newton iterations required to solve the problem".  At that time, Sensor reportedly took 55 Newtons, while CMG took 256, Eclipse 142, and VIP 109.

If anyone is comparing current spe9 run times, the case runs in 1.5 seconds on our 2019 machine with the Intel corei9 9900k chipset and 16 GB memory.

Our experience and feedback from clients is that Sensor usually runs 3 to 10 times faster than other simulators, regardless of problem size. The lower end of the range applies to black oil, the upper end to compositional.  Performance comparisons for this problem are representative of that commonly observed for black oil.

We encourage you to answer the question yourself - Who's Fastest?  For an evaluation, please contact Brian@Coatsengineering.com.

If you want to know what makes Sensor so fast, see the first paragraph of our Why Sensor? page, and Q&A page 1 question 4.

BENCHMARKS

* Intel Xeon Nocona 2.8 GHz, 1MB L2, 800 MHz FSB, running Windows XP

**  Want to see an even faster spe9 run?  The Sensor data given above is untuned, except for the usual formulation and solver selection, and timestep size selections initially and at times that well constraints are changed.  The data file spe9_tol2.dat has an additional entry for solver tolerance, TOLERANCE 2. Run this data in Sensor for a 12% speedup.  Total cpu drops from 8.2 down to 7.2 seconds on our old 2.8 GHz machine, and Sensor is then 4 to 5 times faster than Nexus on this problem!

1. Coats, B.K., Fleming, G.C., Watts, J.W., Shiralkar, G.C., and Rame, M., "A Generalized Wellbore and Surface Facility Model, Fully Coupled to a Reservoir Simulator", SPE Reservoir Evaluation and Engineering, Vol. 7, No. 2, April, 2004.

2. Killough, J.H., "Ninth SPE Comparative Solution Project: A Reexamination of Black-Oil Simulation", SPE 29110, presented at the 13th SPE Symposium on Reservoir Simulation, San Antonio, TX, Feb. 12-15, 1995.