Before MPI: Locality and Kernel Policy in One-Sided Jacobi SVD

This study uses one-sided Jacobi SVD as a laboratory for local design choices in parallel numerical software. It measures runtime, memory footprint, convergence, and reconstruction accuracy while avoiding the mistake of treating MPI wrapper cost as evidence of distributed scalability.

Column-contiguous storage emerges as the dominant choice in this implementation. More elaborate dynamic scheduling reduces some sweeps but does not repay its observation cost. The central lesson is that local memory trajectories and kernel-policy cost models must be sound before scaling across ranks.