Accelerating Simulations of Turbulent Compressible Flows with OpenACC

The field of computational fluid dynamics (CFD) struggles with challenges to resolve computational complexity and cost of the simulations. Executing the operations over large fine meshes at small timesteps requires the algorithms to be designed with parallel efficiency at their core. 

SOD2D, a code developed at Barcelona Supercomputing Center (BSC-CNS) and designed for scale-resolving simulations of compressible turbulent flows in realistic complex cases, used the OpenACC programming model to allow for efficient parallel operations on multi-GPUs. First implemented at the 2023 ALCF INCITE Hackathon, OpenACC was easy to use and brought massive improvements.

The algorithm shows striking performance, where a heavily refined mesh with ~ 48M nodes takes only about 370ms per time-step on a single GPU – a speed that would require several nodes on a supercomputer to reach parity.

Watch the presentation by Lucas Gasparino F. da Silva where he shares how the team behind SOD2D achieved both high accuracy and robustness in flow simulations and industry needs with a remarkable 10X speedup in compute algorithm.