We present LLAMA, a graph storage and analysis system that supports mutability and out-of-memory execution. LLAMA performs comparably to immutable main-memory analysis systems for graphs that fit in memory and significantly outperforms existing out-of-memory analysis systems for graphs that exceed main memory. LLAMA bases its implementation on the compressed sparse row (CSR) representation, which is a read-only representation commonly used for graph analytics. We augment this representation to support mutability and persistence using a novel implementation of multi-versioned array snapshots, making it ideal for applications that receive a steady stream of new data, but need to perform whole-graph analysis on consistent views of the data. We compare LLAMA to state-of-the-art systems on representative graph analysis workloads, showing that LLAMA scales well both out-of-memory and across parallel cores. Our evaluation shows that LLAMA's mutability introduces modest overheads of 3-18% relative to immutable CSR for in-memory execution and that it outperforms state-of-the-art out-of-memory systems in most cases, with a best case improvement of 5x on breadth-first-search.

We present Scalable Host-tree Embeddings for Efficient Partitioning (Sheep), a distributed graph partitioning algorithm capable of handling graphs that far exceed main memory. Sheep produces high quality edge partitions an order of magnitude faster than both state of the art offline (e.g., METIS) and streaming partitioners (e.g., Fennel). Sheep's partitions are independent of the input graph distribution, which means that graph elements can be assigned to processing nodes arbitrarily without affecting the partition quality. Sheep transforms the input graph into a strictly smaller elimination tree via a distributed map-reduce operation. By partitioning this tree, Sheep finds an upper-bounded communication volume partitioning of the original graph. We describe the Sheep algorithm and analyze its space-time requirements, partition quality, and intuitive characteristics and limitations. We compare Sheep to contemporary partitioners and demonstrate that Sheep creates competitive partitions, scales to larger graphs, and has better runtime.

Systems that capture and store data provenance, the record of how an object has arrived at its current state, accumulate historical metadata over time, forming a large graph. Local clustering in these graphs, in which we start with a seed vertex and grow a cluster around it, is of paramount importance because it supports critical provenance applications such as identifying semantically meaningful tasks in an object's history. However, generic graph clustering algorithms are not effective at these tasks. We identify three key properties of provenance graphs and exploit them to justify two new centrality metrics we developed for use in performing local clustering on provenance graphs.

The explosion of graph data in social and biological networks, recommendation systems, provenance databases, etc. makes graph storage and processing of paramount importance. We present a performance introspection framework for graph databases, PIG, which provides both a toolset and methodology for understanding graph database performance. PIG consists of a hierarchical collection of benchmarks that compose to produce performance models; the models provide a way to illuminate the strengths and weaknesses of a particular implementation. The suite has three layers of benchmarks: primitive operations, composite access patterns, and graph algorithms. While the framework could be used to compare different graph database systems, its primary goal is to help explain the observed performance of a particular system. Such introspection allows one to evaluate the degree to which systems exploit their knowledge of graph access patterns. We present both the PIG methodology and infrastructure and then demonstrate its efficacy by analyzing the popular Neo4j and DEX graph databases.