Buffered Streaming Graph Partitioning

TL;DR

This paper introduces a buffered streaming graph partitioning algorithm that improves partition quality on large graphs using minimal memory, bridging the gap between fast streaming and high-quality offline methods.

Contribution

It presents a novel buffered streaming approach that employs multilevel algorithms on model graphs, significantly enhancing partition quality with low memory usage.

Findings

Achieves 75.9% better solutions than Fennel on average.

Removes dependency on number of blocks from runtime.

Faster than Fennel for large number of blocks.

Abstract

Partitioning graphs into blocks of roughly equal size is widely used when processing large graphs. Currently there is a gap in the space of available partitioning algorithms. On the one hand, there are streaming algorithms that have been adopted to partition massive graph data on small machines. In the streaming model, vertices arrive one at a time including their neighborhood and then have to be assigned directly to a block. These algorithms can partition huge graphs quickly with little memory, but they produce partitions with low solution quality. On the other hand, there are offline (shared-memory) multilevel algorithms that produce partitions with high quality but also need a machine with enough memory. We make a first step to close this gap by presenting an algorithm that computes significantly improved partitions of huge graphs using a single machine with little memory in streaming setting. First, we adopt the buffered streaming model which is a more reasonable approach in practice. In this model, a processing element can store a buffer, or batch, of nodes before making assignment decisions. When our algorithm receives a batch of nodes, we build a model graph that represents the nodes of the batch and the already present partition structure. This model enables us to apply multilevel algorithms and in turn compute much higher quality solutions of huge graphs on cheap machines than previously possible. To partition the model, we develop a multilevel algorithm that optimizes an objective function that has previously shown to be effective for the streaming setting. This also removes the dependency on the number of blocks k from the running time compared to the previous state-of-the-art. Overall, our algorithm computes, on average, 75.9% better solutions than Fennel using a very small buffer size. In addition, for large values of k our algorithm becomes faster than Fennel.