Parallel Community Climate Model

#raggedright429#

PCCM2 is a message-passing implementation of the NCAR Community Climate Model 2 (CCM2). In September 1993, PCCM was officially validated with respect to the sequential version of CCM. The SP1 was used extensively in the validation work because its nodes are identical to workstation platforms running the previously validated sequential version. The first validated version of PCCM ran on the SP1. The model is patch decomposed in two horizontal dimensions. Spectral transport of all prognostic variables except moisture is accomplished by parallel FFTs in the zonal dimension and Gaussian quadrature in the meridional dimension, approximating Legendre transforms. The spectral transport mechanism of CCM2 is communication intensive because interchange of data is not confined to nearest neighbor. A semi-Lagrangian transport scheme is used for transport of moisture. Modules that compute atmospheric processes such as convection, radiative transfer, and precipitation are collectively known within the model as <#212#> physics<#212#>. Physics is perfectly parallel in PCCM because there are no horizontal data dependencies; however, physics does present the largest source of inefficiency from load imbalance. PCCM2 was implemented using Chameleon through a compatibility library to PICL, the message-passing package under which the code was originally developed. Before being run on the IBM <#213#>SP1<#213#>, PCCM2 was run on the Intel Touchstone DELTA and Paragon computers. PCCM ran at approximately 650 Mflops on the full <#214#>SP1<#214#> (128 processors) communicating over the EUIH switch interface. Figure #figpccm2use#215> shows the distribution of run time over the three main components of the code: spectral transport, semi-Lagrangian transport, and physics. Interprocessor communication accounted for most of the time spent in the forward and inverse FFTs, the parallel vector-sum (part of the Legendre approximation), and initialization for the semi-Lagrangian transport (SLTINI). The largest computational part of the code is physics, and the effect of load imbalance can be seen as well. The work is being performed under the directed portion of the Department of Energy CHAMMP initiative and is the collaborative effort of Argonne, Oak Ridge National Laboratory, and NCAR. The model is used for climate prediction.

#figure216#
Figure: Distribution of work in a PCCM2 run

%