HPC Scale Engineering Analysis

Scaling CFD workflows to larger HPC resources, local or cloud, requires rethinking data flows

Simply doing “more of the same” is not a strategy for success.

“we were fine when everything was steady state but now LES is killing our file system”

At some point, organizations need to realize that the important knowledge they need is buried in the huge volume files produced by CFD. Solver codes may need a large domain around the article being analyzed by quantities like forces, moments and heat fluxes are usually located on the walls of the article. This can be a few percent of the total stored data size.

New supercomputers have heterogeneous computing architectures and ARM64 is now proving itself. Users need to benchmark and plan end-to-end workflows in order to get the most from these resources.

The CREATE-AV animation shown here was computed on supercomputers in Hawaii, with surface extracts produced in situ. These were transferred over the internet to the engineer’s desktop in the eastern US for post-processing and animation. Without the 1000 fold reduction in file size, the project would not have been possible.

Unsteady LES runs or large parameter sweeps can generate huge and numerous files that become a bottleneck to effective engineering workflows

Is disk space and write/read time limiting the cases I run?

Why do catalyst and libsim use far more memory than my solver?

How can I really use unsteady results for analysis at high timestep frequency?

Are my competitors gaining with new technology that we can’t seem to use?