Accelerating the GFDL Climate Model for the Knights Landing Next-Generation Architecture

The GFDL model is one of the premier climate models that is currently included in the Coupled Model Intercomparison Project Phase 6 (CMIP6). Computational expense is an important practical consideration, and next-generation architectures promise to accelerate software packages. Using Intel’s Vtune Amplifier on a Xeon Phi (codename Knight’s Landing, or KNL) Many Integrated Core machine to perform hotspot and memory access analyses, model code blocks were identified as candidates for rewrites. The code block rewrites that showed significant improvements were collected into a code base. The result of the rewrites leads to a comparison of the original and rewritten code speeds. Collectively, they show a significant speed up of 10-12% and a reduction of the L2 cache miss rate of about 40% on the KNL platform. Most of these improvements come from the rewrite of an aerosol optics routine, with a 15x speedup and an L2 cache miss reduction of over 99%, although a number of other code blocks saw significant improvements as well. Additionally, some practical lessons from this exercise have been learned, particularly where Fortran best practices clash with model design and where different best practices in Fortran come into conflict, as well as considerations that may help the community to resolve these computational issues where they exist.