Abstract
Reproducibility of research results is a fundamental quality criterion in science; thus, computer architecture effects on simulation results must be determined. Here, we investigate whether an ensemble of runs of a regional climate model with the same code on different computer platforms generates the same sequences of similar and dissimilar weather streams when noise is seeded using different initial states of the atmosphere. Both ensembles were produced using a regional climate model named COSMO-CLM5.0 model with ERA-Interim forcing. Divergent phase timing was dependent on the dynamic state of the atmosphere and was not affected by noise seeded by changing computers or initial model state variations. Bitwise reproducibility of numerical results is possible with such models only if everything is fixed (i.e., computer, compiler, chosen options, boundary values, and initial conditions) and the order of mathematical operations is unchanged between program runs; otherwise, at best, statistically identical simulation results can be expected.
Highlights
Reproducibility of research results is a fundamental quality criterion in science; computer architecture effects on simulation results must be determined
The same model setup, i.e., model version, compiler options, boundary conditions and external parameters was used on all the High Performance Computing (HPC) platforms
The first ensemble of six simulations with the same initial conditions was run on different platforms, namely, two platforms at Deutsches Klimarechenzentrum (DKRZ) and one platform at Swiss National Supercomputing Centre (CSCS), LeibnizRechenzentrum (LRZ), Zentralanstalt für Meteorologie und Geodynamik (ZAMG), and Deutscher Wetterdienst (DWD)
Summary
Reproducibility of research results is a fundamental quality criterion in science; computer architecture effects on simulation results must be determined. In the case of limited-area models, the issue was long disregarded because it was falsely assumed that the boundary values would suppress such divergence; this is not the case for regional dynamical models of the atmosphere[4] or the ocean[5] In this case, models can display divergence in phase space in certain periods, while in other time periods, all trajectories may remain close to each other. When a regional atmospheric or oceanic model is initiated with observed states that are a day or more apart, the trajectory of the system will show at a later and potentially much later time, periods of divergence (“intermittent divergence in the phase space”[6]) In this case, the stochasticity is rooted in the very high dimension of the problem and the numerous nonlinear terms. The presence of internal variability is a property of the system
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
