Shogun: A Task Scheduling Framework for Graph Mining Accelerators

Graph mining is an emerging application of great importance to big data analytic. Graph mining algorithms are bottle-necked by both computation complexity and memory access, hence necessitating specialized hardware accelerators to improve the processing efficiency. Current accelerators have extensively exploited task-level and fine-grained parallelism in these algorithms. However, their task scheduling still has room for optimization. They use either breadth-first search, depth-first search or a combination of both, leading to either poor intermediate data locality, low parallelism or inter-depth barriers. In this paper, two key insights on graph mining scheduling are gained, inspired by which a novel task scheduling framework named Shogun is proposed. First, task execution can be out-of-order to eliminate unnecessary barriers and improve PE utilization rate. Second, sacrificing intermediate data locality causes little harm to performance, when a locality monitoring mechanism is adopted to avoid severe locality loss. Hence, Shogun enables adaptive locality-aware out-of-order task scheduling by deploying a task tree to decouple the task generation and execution pipeline stages. Moreover, based on the flexible scheduling design, Shogun further develops accelerator optimizations including task tree splitting for load balance, and search tree merging to explore multiple search trees in parallel on one PE. Experimental results show that Shogun improves performance of a state-of-the-art accelerator by 63% on average with an area overhead of approximately 4%.