Traversing large graphs in realistic settings

Abstract

The notion of graph traversal is of fundamental importance to solving many computational problems. In many modern applications involving graph traversal such as those arising in the domain of social networks, Internet based services, fraud detection in telephone calls etc., the underlying graph is very large and dynamically evolving. This thesis deals with the design and engineering of traversal algorithms for such graphs. We engineer various I/O-efficient Breadth First Search (BFS) algorithms for massive sparse undirected graphs. Our pipelined implementations with low constant factors, together with some heuristics preserving the worst-case guarantees makes BFS viable on massive graphs. We perform an extensive set of experiments to study the effect of various graph properties such as diameter, initial disk layouts, tuning parameters, disk parallelism, cache-obliviousness etc. on the relative performance of these algorithms. We characterize the performance of NAND flash based storage devices, including many solid state disks. We show that despite the similarities between flash memory and RAM (fast random reads) and between flash disk and hard disk (both are block based devices), the algorithms designed in the RAM model or the external memory model do not realize the full potential of the flash memory devices. We also analyze the effect of misalignments, aging, past I/O patterns, etc. on the performance obtained on these devices. We also consider I/O-efficient BFS algorithms for the case when a hard disk and a solid state disk are used together. We present a simple algorithm which maintains the topological order of a directed acyclic graph with n nodes under an online edge insertion sequence in O(n2.75) time, independent of the number m of edges inserted. For dense DAGs, this is an improvement over the previous best result of O(min{m 3 2 logn,m 3 2 +n2 logn}). While our analysis holds only for the incremental setting, our algorithm itself is fully dynamic.