High-Performance with an In-GPU Graph Database Cache

Although graph databases have great potential to directly represent huge and complex network structures, such as social and logistics networks, they currently do not perform well enough. To achieve a two-orders-of-magnitude performance improvement on graph database queries, the authors propose caching the graph database in distributed GPUs connected with a 10-Gbit Ethernet (10GbE) network. Property read, traverse, and graph search queries are accelerated by 25.9, 252.4, and 383.2 times, respectively, compared to those using Neo4j for a huge graph consisting of 3.2 million nodes whose degree is 100.