High Performance Concurrency Control and Commit Protocols in OLTP Databases

Abstract

Depending on the required scale, modern data-orientated applications are built on top of either many-core or distributed OLTP databases. In both architectures, database concurrency control is a performance critical component. Additionally, in distributed OLTP databases, where a transaction can span across multiple data partitions, distributed atomic commitment is a necessity; unless the latter is designed judiciously, it can significantly degrade performance. This thesis explores both facets to address the inefficiencies and limitations of existing work. The wait-hit protocol is a serializable concurrency control protocol that is designed to offer high performance across all scale points. This thesis also investigates mixed serialization graph testing by applying the recently revived graph-based approach to concurrency control, which minimizes unnecessary aborts, to cater for transactions running at weaker isolation levels, a phenomenon common in practice. Protocols are developed for ensuring reciprocal consistency and edge-order consistency in distributed graph databases, guarantees unique to graph databases which without sufficient concurrency control can be violated causing irreparable data corruption. Finally, we demonstrate that epoch-based distributed databases can amortize atomic commitment costs. In addition to developing an analytical model for choosing the optimal epoch size in such databases, this thesis presents epoch-based multi-commit which can reduce wasted work when database nodes fail. Each protocol has been subjected to extensive performance evaluation either through simulations or implementation. Our experiments show that each protocol offers a marked improvement over the current state-of-the-art.