Parallelizing legacy AFLR mesh generator hits 11x speedup but faces limits

A team led by Kevin Garner from Mississippi State University documented their attempt to parallelize AFLR, a widely used legacy mesh generation tool for computational simulations. They treated AFLR as a black box, splitting the domain into subdomains and running the serial code on each in parallel, with a runtime system for load balancing. On a mesh refinement task, the parallel method achieved roughly 11 times speedup using 16 CPU cores compared to the serial version. Output mesh quality was good, but not identical—the parallel method produced a different mesh volume due to boundary constraints imposed by the black-box interface. Satisfying those constraints added overhead and prevented full reproducibility.

More critically, the team found that achieving full stability (matching serial quality) and better scalability would require non-trivial modifications to AFLR's internals. The black-box approach introduced fundamental limits: each subdomain needed boundary data that changed the local refinement process. The authors conclude that parallelizing legacy black-box codes like AFLR may not be practical, and instead advocate designing new codes from scratch with parallelism in mind. Their work serves as a cautionary tale for high-performance computing efforts that rely on legacy solvers, emphasizing that brute-force parallelization of opaque software often hits diminishing returns.