PNNL

Abstract

A test procedure is proposed for identifying numerically significant solution changes in atmospheric models that solve the partial differential equations of fluid dynamics. The test issues a “fail” signal when any code modifications or computing environment changes lead to solution differences that exceed the known time step sensitivity of the reference model. It is demonstrated using the Community Atmosphere Model version 5.3 (CAM5.3) that the proposed procedure can correctly distinguish rounding-level changes in the solutions from impacts of compiler optimization or parameter perturbations that are known to cause non-negligible differences in the simulated climate. The short simulation length implies low computational cost, and makes the test useful for debugging. The independence between ensemble members allows for parallel execution of all simulations thus facilitates fast turnaround. The version 1.0 implementation described in the present paper uses 12-member 5-minute simulations. The computational cost of producing the reference results is close to a 4-month simulation conducted using the default model time step, and the cost of testing a new code or computing environment is close to a 1-month simulation conducted using the default model time step. Tests carried out at the leadership computing facilities indicate it is possible to finish the integration of all ensemble members within a few minutes. The new method is simple to implement since it does not require any code modifications. We expect the same methodology can also be used for any geophysical models to which the concept of time step convergence is applicable.