Revival of Vector Processors in Earth System Modelling

Panagiotis Adamidis (DKRZ)

Nowadays, climate simulations are of high societal relevance. In order to reduce significantly the uncertainty of climate change projections, it is crucial to understand several physical processes, like forming of clouds and precipitation. Within the HD(CP)2 project an LES model based on ICON (ICOsahedral Non hydrostatic GCM) has been developed. The ICON-LEM model aims at resolving cloud and precipitation processes using grids with a resolution of 10000x10000x400 grid elements and a grid spacing of 100m. Such simulations are computationally very intensive. Although, modern high performance computing platforms can deliver a peak performance in the range of Petaflop/s, ICON is not able to exploit this power.

Current HPC systems are massively parallel computers and achieving good scalability is vital in order to be able to use systems consisting of hundreds of thousands of cores efficiently. The two major bottlenecks are fast parallel I/O and moving data between the levels of the memory hierarchy inside cores/nodes or among nodes. Avoiding communication by increasing temporal and spatial locality is a key issue.

The NEC Aurora TSUBASA System promises high sustained performance by combining high floating point performance of a newly designed vector processor with extremely high memory bandwidth. The presentation will show first results from our tests with earth system models on the NEC Aurora test system.