Acceleration of Document-Oriented Databases Using GPUs

Document-oriented databases are databases that can store documents in a schema-less manner and perform search queries for the documents. They can be used for web applications and online games that require high scalability and rich functions. The mainly function of document-oriented databases is a string search for documents. Computation cost of a database search is proportional to the number of documents and thus it is very high for large-sized documents. Although database indexes are used in document-oriented databases to reduce the computation cost, they cannot be applied for all the string search queries, such as regular expression matching queries. In this paper, we propose DDB Cache (Document-oriented DataBase Cache) to accelerate these queries with GPUs. The DDB Cache structure is suitable for GPU-based string processing and thus we can improve the string search query performance without using database indexes. In addition, we propose a DDB Cache partition method using a hash function to partition and distribute DDB Cache to multiple GPUs, so that we can improve the horizontal scalability of GPU processing with DDB Cache. We implemented DDB Cache for MongoDB. Experimental results show that our approach significantly outperforms the original MongoDB in regular expression search queires. The results also show that, when we increase the number of GPUs from one to three in our approach, the regular expression matching query throughput is improved by 2.7 times, resulting in a favorable horizontal scalability.