Mixed-precision for Linear Solvers in Global Geophysical Flows.

摘要: Semi-implicit time-stepping schemes for atmosphere and ocean models require elliptic solvers that work efficiently on modern supercomputers. This paper reports our study of the potential computational savings when using mixed precision arithmetic in solvers. The essential components a representative solver are run at levels as low half (16 bits), accompanied with detailed evaluation impact reduced convergence solution quality. A inquiry into requires model configuration is meaningful but cheaper to easier evaluate than full atmosphere/ocean models. therefore conducted context novel semi-implicit shallow-water sphere, purposely designed mimic numerical intricacies all-scale weather climate (W&C) stability independent celerity all wave motions. governing algorithm based non-oscillatory MPDATA methods geophysical flows, whereas resulting problem employs strongly preconditioned non-symmetric Krylov-subspace GCR, proven advanced atmospheric applications. classical longitude/latitude grid deliberately chosen retain stiffness global W&C posed thin spherical shells well better understand performance reduced-precision vicinity singularities. Precision reduction done software level, an emulator. experiments established dynamical-core test-cases, like Rossby-Haurwitz number 4 zonal orographic flow. The shows selected key solver, most prominently preconditioning, can be performed level precision.