Oak Ridge National Laboratory
Measuring Up on the Petascale
(Page 2 of 3)
Bridging Worlds
Climate modeling is one area expected to benefit from the great leap forward petascale computing offers.
There is a concerted effort to add the full carbon cycle to climate models, Kothe explains. Previous research has suggested ecosystems respond to increasing atmospheric carbon dioxide by taking up more carbon, but paradoxically also respond to increasing temperature by releasing carbon. Scientists would like to compare these responses in a full climate simulation, but it has been too expensive in terms of computer time and cost to add a fully functional carbon cycle to the models. Instead, models specified the amount of carbon dioxide in the atmosphere as a fixed input.
Snapshot of simulated time evolution of carbon dioxide concentration originating from land, around 1900 Click image for larger version and more information |
The massive project, called “A scalable and extensible Earth system model for climate change science,” includes input from several national laboratories, academic programs and DOE’s Atmospheric Science Program and Terrestrial Carbon Program, among others. Its goal is nothing less than transforming an existing global climate model, the Community Climate System Model (CCSM), to create one that fully simulates coupling between the physical, chemical, and biogeochemical processes in the climate system.
John Drake heads the project’s group at ORNL’s Computer Science and Mathematics Division, and Peter Gent of the National Center for Atmospheric Research is chairman of the group’s scientific steering committee. A group at Pacific Northwest National Laboratory contributes the component that calculates the chemistry of aerosols. Lawrence Livermore National Laboratory provides the chemistry component. Groups at Argonne National Laboratory work on coupling all of the components. The ORNL Leadership Computing team is contributing personnel from NCCS’s Scientific Computing Group, which works directly with users, their software, and data, to get all the components ready for the jump to petascale computing.
“The people in our Scientific Computing Group have a lot of fingertip knowledge, meaning their hands are on the keyboard a lot in terms of writing code, and so they know the latest practices that can help these teams evolve their codes and algorithms to be more productive and more long-lived,” Kothe says.
A background in physics and experience in high-performance computing allows Trey White, a research computer scientist at NCCS, to contribute to several projects. Presently he works on the Parallel Ocean Program (POP), a component developed at Los Alamos National Laboratory for the CCSM. He assists Pat Worley, a computer scientist in the climate group, to streamline POP and improve its scalability for petascale computing.
Worley, who has a long history of POP performance analysis, found a sluggish solver was one thing limiting its scalability. The code must find the sum of a value across all the processors, a step called an “all-reduce,” which significantly slows the parallel operation. White is looking for a way to reduce the need for all-reduce queries by improving what’s called the preconditioner. Meanwhile, Worley is making the entire operation faster by arranging the processes more efficiently.
« Previous 1 | 2 | 3 | References | Sidebar | Print Next »
