Objective: To develop scalable parallel computational algorithms for the main computational tasks in climate models. These tasks include the solution of the equations of motion for the ocean and the atmosphere and the parameterizations of subgrid-scale processes. The parallel algorithms need to be scalable to at least several hundred processors to achieve the desired performance on future massively parallel processors (MPP's).
Approach: Our approach has been to develop modular and generic parallel codes of a representative sample of the computational tasks encountered in climate models. We were thus able to test and compare variety of machines and architectures. We felt that this would be more productive than concentrating on a single machine and expending the effort required to bring up an entire climate model. We also felt that in this way we could be of greater service to the HPCC program by providing codes that would be adaptable to different architectures and could be used as benchmarks for the evaluation of testbed hardware and software, while advancing our own development of scalable models.
Accomplishments: We have produced parallel versions of most of the major computational modules of the atmospheric model, the computational core of the land surface model, and a simplified version of the entire ocean model. We have used these to benchmark a variety of machines, including both shared- and distributed-memory architectures. In all cases, we have overcome most of the major scaling problems on at least one machine and have achieved performances of 5 to 15 GFLOPS on message passing codes and good scaling on communication-intensive codes out to 512 processors. Further, we have done this while maintaining generic and flexible codes capable of running in both message passing or data parallel modes, on vector or cache-based processors, and with distributed- or shared-memory.
Significance: The codes developed during this project should allow us to implement a complete parallel climate model capable of 40 GFLOPS on the next generation of MPP's.
Point of Contact:
Max Suarez
NASA/Goddard Space Flight Center
suarez@maxs.gsfc.nasa.gov
301 286-7373
URL: http://pong.gsfc.nasa.gov/hpcc/