PDG: A Prefetcher for Dynamic Graph Updating

Dynamic graphs can be utilized to model manyreal-world applications like social media analysis in whichthe connections and entities evolve continuously. Hence, theprocessing of dynamic graphs is gaining increasing popularity.However, prior dynamic graph processing systems mainly focuson the optimization of graph analytics but overlook graphupdating which manages the evolving graph structure andpresents a unified view to graph analytics. Since graph updatingoperates on evolving graphs and involves a large number ofirregular memory accesses, it poses a substantial influence onthe performance of dynamic graph processing systems. In thiswork, we observe that graph updating is mainly bottlenecked bya frequent indirect memory access pattern & lowast;(& lowast;(B[A[i]]+offset)).The pattern is inherent to the typical graph updating from theincoming edge stream to the base data store organized witheither an adjacent list or a compressed sparse row. With thisobservation, we propose a novel prefetcher for dynamic graphupdating abbreviated as PDG. PDG is a lightweight pipelinedinstruction-based prefetcher specialized for graph updating andit is also compatible with the irregular memory access patternB[A[i]] widely used in graph analytics. In addition, it leveragesa monitor of the instruction queue to decide the appropriatetiming of prefetching to make the best use of the cache.According to our experiments, PDG achieves 1.60x,1.26x,and1.30xperformance speedup compared to three representativeprefetchers, respectively, with negligible hardware overhead ingraph updating