Parallel Array-Based Single- and Multi-Source Breadth First Searches on Large Dense Graphs

One of the fundamental algorithms in analytical graph databases is breadth-first search (BFS). It is the basis of reachability queries, centrality computations, neighborhood enumeration, and many other commonly-used algorithms. We take the idea of purely array-based BFSs introduced in the sequential multi-source MS-BFS algorithm and extend this approach tomulti-threaded singleandmulti-source BFSs. Replacing the typically used queues with fixed-sized arrays, we eliminate major points of contention which other BFS algorithms experience. To ensure equal work distribution between threads, we co-optimize work stealing parallelization with a novel vertex labeling. Our BFS algorithms have excellent scaling behavior and take advantage of multi-core NUMA architectures. We evaluate our proposed algorithms using real-world and synthetic graphs with up to 68 billion edges. Our evaluation shows that the proposed multi-threaded singleand multisource algorithms scale well and provide significantly better performance than other state-of-the-art BFS algorithms.