Speedup Performance of 3DSL Parallel
Want to know approximately how much faster 3DSL will be if you run on multiple cores? Unfortunately it is not as straightforward as doubling CPU's to double performance.
The maximum speedup is related to the overall fraction of parallelized runtime code. Depending on the physics of your model, the number of gridblocks, the number of timesteps and timestep size, determines the fraction of parallelized code. See the chart below to give you a guide as to the speedup performance of 3DSL as you add cores. Note that the greatest incremental gain in speedup is for a low number of CPU's. Regardless of model type, you can expect about a 1.6X speedup in run time just by running 3DSL on a dual-core system. Streamsim is continually working to parallelize additional regions of 3DSL meaning that speedup performance will continue to improve with future releases of 3DSL.
 |
Speedup of 3DSL parallel for various simulation models as number of CPU's increases. |
- SPE10-inc. Incompressible 2-phase. 1.1 million active grid cells, 25 timesteps over 2000 days.
- SPE10-comp. Compressible deadoil. 1.1 million active gridcells, 27 timesteps over 2000 days.
- Forties. Incompressible 2-phase. 1.5 million active grid cells, 235 wells, 82 timesteps over 34 years.
- JudyCreek. Incompressible 3-phase FCM. 623,000 active grid cells, 300 wells, 46 timesteps over 40 years.
- SA-yearly. Compressible deadoil, 970,000 active grid cells, 25 wells, aquifer, 24 timesteps over 24 years. SA-monthly is the same run but with 288 timesteps.
- ME-Dualphi. Incompressible, 706,000 active grid cells, dualporosity, 145 producers, 60 timesteps over 34 years.
All results were for a 4-core dual CPU 5460 Xeon, Windows Vista, 3DSL v4.10 (2010/06/15).
For further details, see SPE paper 118684 |
