Productive Performance Engineering for Weather and Climate Modeling with Python

Explore productive performance engineering techniques for weather and climate modeling using Python in this conference talk from Supercomputing '22. Dive into the optimization of the Finite Volume Cubed-Sphere Dynamical Core (FV3) through a declarative Python-embedded stencil domain-specific language and data-centric optimization. Learn about a semi-automated workflow for analyzing and optimizing weather and climate applications, utilizing both local and full-program optimization, as well as user-guided fine-tuning. Discover how the novel transfer tuning approach prunes the infeasible global optimization space by automatically utilizing repeating code motifs. Examine the FV3GFS Model, the Pace Project, and the shift of scientific computing towards Python. Gain insights into GridTools for Python (GT4Py) and Dace Overview. Explore the characterization of the optimization space, evaluated systems, memory bounds, and representative vertical and horizontal stencils. Conclude with an understanding of weak scaling and how these techniques achieved speedups of up to 3.92x over tuned production implementation on the Piz Daint supercomputer, scaling to 2,400 GPUs.

Productive Performance Engineering for Weather and Climate Modeling with Python
The FV3GFS Model
The Pace Project
Scientific Computing is Moving to Python
GridTools for Python (GT4PY)
Dace Overview
Characterizing the optimization space
Evaluated Systems
Memory Bounds
Representative Vertical Solver
Representative Horizontal Stencil
Weak Scaling