Accelerating Spanning-Centrality Approximation with Sampling-Aware Vectorized Random Walks
Spanning centrality quantifies an edge's importance through its share of a graph's spanning trees, and approximating it for all edges (AESC) requires processing large volumes of irregular random walks under a prescribed error bound. Its applications include phylogenetic analysis and network robustness assessment. We focus on this random-walk processing bottleneck and propose Sampling-Aware Bouquet Arrangement (SABA), a CPU sampling and scheduling scheme that orders walks initialized on a uniform grid so that adjacent walks execute together in SIMD-width groups, called bouquets. For fixed per-walk random inputs, the ordering preserves the multiset of sampled walks while improving locality during neighbor access. On synthetic benchmarks, SABA achieves geometric-mean speedups of 6.95x and 7.41x over traditional random-walk generation for 2,048 and 16,384 walks per vertex, respectively, and up to approximately 49x in the latter setting. Within AESC, the complete 16-thread SABA implementation is 58.23x faster than the original single-threaded TGT+ implementation on average.