An ensemble analysis of forecast errors related to floating point performance

Abstract

The dynamical core of the Mesoscale Compressible Community (MC2) model is described. Ensemble forecast techniques for high-resolution mesoscale simulations are applied to assess the impact of floating point optimization, mathematics libraries, and processor configuration on forecast accuracy. It is shown that the iterative solver in the dynamical core is most sensitive to processor configuration, but it also shows weak sensitivity to the usage of fast mathematics libraries and floating point instruction reordering. Semi-implicit pressure solver errors are amplified in the physical parameterization package, which is sensitive to small pressure differences and feeds back to the dynamical solution. In this case, local rms spreads are around 1°C in temperature by the end of a 42-h forecast. It is concluded that careful validation is required when changing computing platforms or introducing fast mathematics libraries.