|Contact Us | Visit Us | UCAR People Search||Numerics| Assimilation| Turbulence| Statistics|
|[an error occurred while processing this directive]||
High-Order Methods Modeling Environment
HOMME Team: John Dennis, Jim Edwards, Rich Loft, Ram Nair, Amik St-Cyr, Mark Taylor, Steve Thomas, Henry Tufo
NCAR researchers, funded in part by the Department of Energy's Climate Change Prediction Program, have built a scalable and efficient spectral-element-based atmospheric dynamical core using the Computer Science Section and the Computational Mathematics Group High Order Method Modeling Environment (HOMME). In order for this to be a useful tool for atmospheric scientists it is necessary to couple this core to physics packages employed by the community.
The model is written in F90 and has three components: dynamics, physics and a physics/dynamics coupler. The dynamics has been run on the BG/L systems at Watson and Rochester on up to 7776 processors using one processor per node and only using one of the floating-point pipelines. The peak performance expected from a Blue Gene/L processor for the runs is then 1.4 Gflops/s. The sustained Mflops per second per processor is shown below for meshes with 1536 (16x16x6), 1944 (18x18x6), and 7776 (36x36x6) elements on the sphere and with either 20 or 40 layers in the vertical direction. The total amount of work (total flop count to solve the system) was kept constant while the processor count was increased (strong scaling), until, in the 1944 and 7776 mesh runs, there was only one element per processor. The flat part of the graph shows the scaling region for each problem size, which would be the region where one would make the most effective use of the BG/L processor. The average sustained performance in the scaling region for the Dry Held-Suarez code is ~200-250 MF/s/processor (out to 7776 processors) while for the Moist Held-Suarez code it is ~300-400 MF/s/processor (out to 1944 processors) out of a peak of 1400 MF/s/processor.
To join the discussion list or browse the Homme Archives please visit: