SIMD-X: Programming and Processing of Graph Algorithms on GPUs
Talkat USENIX ATC '19 · on Thursday, 2019-07-11· at 11:15· USENIX ATC Track II: Grand Ballroom VII–IX
summary
Presents a GPU graph framework that exposes an Active-Compute-Combine (ACC) API so algorithms are written in tens of lines, uses just-in-time task management to filter inactive vertices and balance workload, and applies push‑pull kernel fusion plus a deadlock-free global barrier to slash kernel counts. Reports real speedups: about 3x vs Gunrock, 6x vs Galois, 24x vs CuSha, 3x vs Ligra on social and web graphs. Notes overhead on uniform-degree graphs and limits for high‑dimensional ML-style vertex vectors or workflows needing more complex control flow.
AI-written summary · about