LAUSR

Abstract. Based on the fifth phase of the Coupled Model Intercomparison Project (CMIP5)-generation previous Institut Pierre Simon Laplace (IPSL) Earth system model, we designed a new version, IPSL-CM5A2, aiming at running multi-millennial simulations typical of deep-time paleoclimate studies. Three priorities were followed during the setup of the model: (1) improving the overall model computing performance, (2) overcoming a persistent cold bias depicted in the previous model generation and (3) making the model able to handle the specific continental configurations of the geological past. These developments include the integration of hybrid parallelization Message Passing Interface – Open Multi-Processing (MPI-OpenMP) in the atmospheric model of the Laboratoire de Meteorologie Dynamique (LMDZ), the use of a new library to perform parallel asynchronous input/output by using computing cores as “I/O servers” and the use of a parallel coupling library between the ocean and the atmospheric components. The model, which runs with an atmospheric resolution of 3.75 ∘ × 1.875 ∘ and 2 to 0.5 ∘ in the ocean, can now simulate ∼100  years per day, opening new possibilities towards the production of multi-millennial simulations with a full Earth system model. The tuning strategy employed to overcome a persistent cold bias is detailed. The confrontation of a historical simulation to climatological observations shows overall improved ocean meridional overturning circulation, marine productivity and latitudinal position of zonal wind patterns. We also present the numerous steps required to run IPSL-CM5A2 for deep-time paleoclimates through a preliminary case study for the Cretaceous. Namely, specific work on the ocean model grid was required to run the model for specific continental configurations in which continents are relocated according to past paleogeographic reconstructions. By briefly discussing the spin-up of such a simulation, we elaborate on the requirements and challenges awaiting paleoclimate modeling in the next years, namely finding the best trade-off between the level of description of the processes and the computing cost on supercomputers.