SMS: A Directive-Based Parallelization Approach for Shared and Distributed Memory High Performance Computers

Dan Schaffer
OAR/FSL Aviation Division

The Forecast Systems Laboratory (FSL) has developed the Scalable Modeling System (SMS); a directive-based tool for parallelizing weather and ocean models. The user inserts SMS directives in the form of comments into existing FORTRAN code. SMS translates the code and directives into a parallel version that runs efficiently on both distributed and shared memory machines. Included are directives that address nested and coupled ocean/atmosphere models. SMS provides directives that greatly simplify the debugging process as well as a mechanism for parallelizing a model piecemeal. Although targeted toward finite difference geophysical codes, the SMS directives are general enough to support a wide range of applications. SMS has been applied to models at FSL, the Pacific Marine Environmental Laboratory (PMEL), the Environmental Research Laboratory (ERL), the Aeronomy Laborary (AL), the Los Alamos National Laboratory (LANL) and other non-government institutions.

Here we describe the SMS tool and show how it has been used to parallelize the Rapid Update Cycle (RUC) model. Scalability of SMS RUC is analyzed for the NCEP SP2 and FSL's Alpha Linux cluster. The SMS version of the RUC model is currently being run in production on the NCEP SP2. We also examine the performance of an SMS version of the Eta model. SMS Eta performance has been compared to that of the hand-coded MPI version used at NCEP and found to be equivalent.

BIO - Dan Schaffer BA in Math and Computer Science from UC Berkeley 1983 MS in Meteorology from Univ of MD, College Park 1993. Worked for NASA/Goddard parallelizing Max Suarez's atmospheric model from 1994-1998 Worked for FSL in the Advanced Computing Group 1998-now.

Conference Center - Paper Wednesday - 10:30 - 11:00 A.M.