Computing shortest path distance is a fundamental primitive in many graph applications. On graphs that do not fit in the main memory of the computing device, computing such distances requires hours to months even with the best I/O-efficient shortest path implementations. For applications requiring many such shortest path distances, one would ideally like to preprocess the input graph into a space-efficient data structure I/Oe efficiently, such that the distance queries can be answered with a small additive distortion using only O(1) I/Os. Furthermore, in a batch setting, one would like to answer O(n) such distance queries in ~O (n=B) I/Os. In this paper, we focus on engineering an I/O-efficient distance oracle for large graphs that model real-world interactions. Our engineered oracle (i) preprocesses graphs with multi-billion edges in less than an hour using a single core of a typical PC, (ii) answers online shortest path queries in milliseconds using a SSD, (iii) answers batched shortest path queries using HDDs with an average time per query of a few microseconds, (iv) results in a highly accurate shortest path estimate and (v) uses space linear in the number of nodes. Our implementation creates small oracle labels (i.e., they can still be kept in internal memory for rather large graphs) but also efficiently handles the case when both the graph and these labels have to reside on external storage. Dynamic settings where new edges are continuously inserted into the graph are efficiently supported, too.
展开▼
