Abstract
The distribution of a climate model across homogeneous and heterogeneous computer environments with nodes that can reside at geographically different locations is investigated. This scientific application consists of an atmospheric general circulation model (AGCM) coupled to an oceanic general circulation model (OGCM).
Three levels of code decomposition are considered to achieve a high degree of parallelism and to mask communication with computation. First, the domains of both the gridpoint AGCM and OGCM are divided into subdomains for which calculations an carded out concurrently (domain decomposition). Second, the model is decomposed based on the diversity of tasks performed by its major components (task decompositions). Three such components are identified: (a) AGCM/physics which computes the effects on the grid-scale flow of subgrid-scale processes such as convection and turbulent mixing; (b) AGCM/dynamics, which computes the evolution of the flow governed by the primitive equations; and (c) the OGCM. Task decomposition allows the AGCM/dynamics and OGCM calculations to be carried out concurrently. Last, computation and communication are organized in such a way that the exchange of data between different tasks is carded out in subdomains of the model domain (110 decomposition). In a dedicated computer network environment, the wall-clock time required by the resulting distributed application is reduced to that for the AGCMJ physics, with the other two components and interprocess communications running in parallel.
The network bandwidth requirements for the distributed application are analyzed. It is assumed that the wall-clock time required to run the AGCM/physics for the model atmosphere in a dedicated computer environment is fixed at a value corresponding to high network efficiency. The analysis shows that, for computer environments based an nodes equivalent to the Intel Touchstone Delta, a bandwidth approaching that of the Gigabit Network is required for an efficient operation of the distributed application with model resolution double that used in current studies of the climate system if output is visualized in real time.
It is argued that distribution of a climate model based on domain, task, and 110 decomposition has the potential for significant and eventually superlinear speedup in model execution, which will facilitate performance of the long integrations required by climate studies.