MaiterStore: A Hot-Aware, High-Performance Key-Value Store for Graph Processing

Recently, many cloud-based graph computation frameworks are proposed, such as Pregel, GraphLab and Maiter. Most of them exploit the in-memory storage to obtain fast random access which is required for many graph computation. However, the exponential growth in the scale of large graphs and the limitation of the capacity of main memory pose great challenges to these systems on their scalability. In this work, we present a high-performance key-value storage system, called MaiterStore, which addresses the scalability challenge by using solid state drives (SSDs). We treat SSDs as an extension of memory and optimize the data structures for fast query of the large graphs on SSDs. Furthermore, observing that hot-spot property and skewed power-law degree distribution are widely existed in real graphs, we propose a hot-aware caching (HAC) policy to effectively manage the hot vertices (frequently accessed vertices). HAC can conduce to the substantial acceleration of the graph iterative execution. We evaluate MaiterStore through extensive experiments on real large graphs and validate the high performance of our system as the graph storage.