Workshop: High Performance Computing for Environmental and Earth Sciences (HPC4EES)
Authors: Lorenzo Breschi and Flavio Vella (University of Trento)
Abstract: Operational pseudospectral atmospheric models such as the ECMWF Integrated Forecasting System (IFS) run today almost exclusively on CPUs; GPU ports are under active development but not yet used in production. These models rely on the Spherical Harmonic Transform (SHT) to compute spatial derivatives and to alternate between spectral and grid-space representations. Each time-step requires forward and inverse SHTs, and both passes depend on global pencil transposition: row-wise/column-wise all-to-all collectives that redistribute multi- dimensional arrays across compute nodes. While GPU floating- point throughput has grown rapidly, inter-node network band- width has not kept pace, and at large node counts these global collectives dominate wall-clock time. As an alternative to float16 truncation during communication, we investigate GPU-resident lossy compression, a capability not yet available to any operational CPU-bound pseudospectral model. Using representative fields from the DYAMOND high-resolution operational dataset as input, and combining measured GPU compression throughput with SimGrid network simulation, we show that ZFP at 16 bits per value (rate-16), about the same storage budget as float16, matches the communication-time reduction of float16 truncation while delivering approximately 1600× lower mean relative error. ZFP at 8 bits per value (rate-8) achieves approximately 1.93× the speedup of float16 while retaining 4× lower mean relative error. We additionally corroborate the simulated communication-time trend with real measurements on the CPU partition of the CRESCO8 cluster.
Back to High Performance Computing for Environmental and Earth Sciences (HPC4EES) Archive Listing
Back to Full Workshop Archive Listing