Cache-Guided Scheduling: Exploiting Caches to Maximize Locality in Graph Processing

Graph processing algorithms are currently boŠlenecked by the limited bandwidth and long latency of main memory accesses. Onchip caches are of liŠle help when processing large graphs because their irregular structure leads to seemingly random memory references. However, most real-world graphs o‚er signi€cant potential locality—it is just hard to predict ahead of time. In practice, graphs have well-connected regions where relatively few vertices share edges with many common neighbors. If these vertices were processed together, graph algorithms would enjoy signi€cant reuse. But €nding this cache-friendly schedule is hard from the processor side, which is oblivious to cache contents. Our insight is that the cache knows exactly which vertices are cached at any given time, so it is in an ideal position to €nd a schedule that maximizes locality. We present our ongoing work on CacheGuided Scheduling (CGS), a technique that exploits this insight by adding a specialized engine to the last-level cache that dynamically €nds a schedule that minimizes cache misses. We present a limit study of CGS through two idealized implementations. Œis limit study reveals CGS’s large potential: CGS reduces memory accesses by 5.8× gmean on a set of large graphs. Œough promising, CGS would incur high overheads if implemented this way. We discuss several paths towards a practical implementation.