Efficient Commit Processing in High-Availability Main-Memory Databases

Database systems are currently being used as backbone to thousands of applications. Some of these have very high demands for high availability and fast real-time responses. Examples include telecommunication systems, trading systems, online gaming and sensor networks. Disk-based databases are too slow to facilitate the short response time requirements of these types of systems. Main-memory databases, however, provide short response times and, if replicated, the required level of availability. Still, the execution of transactions should be optimized to support as many transactions per second as possible while keeping the response times at acceptable levels. The focus of this work is how to improve the performance of main-memory primary-backup systems by looking at the execution of transactions and commit protocols. Furthermore, system resilience to failures is addressed. We have examined existing solutions both for distributed environments in general and main-memory primary-backup systems in particular. The main contribution of this thesis deals with protocols and several optimizations for committing transactions in replicated main-memory databases. The protocols reduce the overhead by sending messages in a circular fashion. In addition, the results of an extensive investigation of existing commit protocols and optimizations for them are presented as related work. Furthermore, the protocols are extended to multiple copies and an approach to handle failures in non-deterministic environments. This improves the resilience to failures. The suggested protocols and optimizations have been evaluated. The results show a 50% increase in throughput and a reduction in transaction response time are achieveable compared to state-of-the-art protocols. Throughput for single-tuple transactions can be further improved by 80% – 130%, depending on the ratio between reads and writes. Hence, the work clearly has merit for distributed databases with high availability and short real-time response requirements.