IMPROVEMENT OF SMVGEAR-II ON VECTOR AND SCALAR MACHINES THROUGH ABSOLUTE ERROR TOLERANCE CONTROL

The computer speed of SMVGEAR II was improved markedly on scalar and v ector machines with relatively little loss in accuracy. The improvemen t was due to a method of frequently recalculating the absolute error t olerance instead of keeping it constant for a given set of chemistry. To test the effects of the modification on speed and accuracy, three E ulerian model simulations were performed on two vector machines, the G ray C-90 and Gray J-916, and a scalar machine, the Silicon Graphics Or igin 2000. The first simulation was an urban air pollution case in whi ch gas chemistry, dynamical meteorology, radiation, and species transp ort were solved together. The second and third simulations were global cases in which chemistry, dynamical meteorology, radiation, and speci es transport were solved together. For the urban case, the new algorit hm reduced SMVGEAR II's computer time by 50-67%, depending on the mach ine used. For the global cases, chemistry time was reduced by 14-44%. In all simulations, normalized gross errors were less than 1%, and the time required to solve chemistry dropped to near or below 50% of the total model simulation time. Thus, gas chemistry may no longer be the computationally-limiting process in many atmospheric models. The simul ations show the usefulness of SMVGEAR II on a scalar workstation with sufficient memory bandwidth. (C) 1998 Elsevier Science Ltd. All rights reserved.