TurboGraph++: A Scalable and Fast Graph Analytics System

Abstract

Existing distributed graph analytics systems are categorized into two main groups: those that focus on efficiency with a risk of out-of-memory error and those that focus on scale-up with a fixed memory budget and a sacrifice in performance. While the former group keeps a partitioned graph resident in memory of each machine and uses an in-memory processing technique, the latter stores the partitioned graph in external memory of each machine and exploits a streaming processing technique. Gemini and Chaos are the state-of-the-art distributed graph systems in each group, respectively.

We present TurboGraph++, a scalable and fast graph analytics system which efficiently processes large graphs by exploiting external memory for scale-up without compromising efficiency. First, TurboGraph++ provides a new graph processing abstraction for efficiently supporting neighborhood analytics that requires processing multi-hop neighborhoods of vertices, such as triangle counting and local clustering coefficient computation, with a fixed memory budget. Second, TurboGraph++ provides a balanced and buffer-aware partitioning scheme for ensuring balanced workloads across machines with reasonable cost. Lastly, TurboGraph++ leverages three-level parallel and overlapping processing for fully utilizing three hardware resources, CPU, disk, and network, in a cluster. Extensive experiments show that TurboGraph++ is designed to scale well to very large graphs, like Chaos, while its performance is comparable to Gemini.