Partial Replication on Transactional Memory Systems

Nowadays, transactional in-memory distributed storage systems are widely used as a mean to increase the performance of applications that need to access frequently large amount of shared data. In this context, data replication has two main advantages: it supports load balancing and fault-tolerance. However, these advantages need to be weighted against the costs of replications: namely memory consumption and coordination costs. This motivates the use of partial replication. In order to maximize performance and scalability, these platforms typically rely on weakly consistent partial replication mechanisms, sacrificing consistency and ensuring guarantees that are weaker than classic 1-Copy Serializability. This thesis addresses the problem of supporting partial replication in transactional in-memory distributed storage systems. Although solutions have been proposed for partial replication in database management systems, there are significant differences between the two type of systems. Namely, transactional inmemory distributed storage systems avoid many of the costs involved in the management of database transactions, which amplifies the costs of replica maintenance. Therefore, the thesis aims at assessing the efficiency of partial replication in the context of transactional in-memory distributed storage systems. The thesis presents the following contributions: proposes a set of algorithms to support partial replication in transactional in-memory distributed storage systems; and presents an experimental evaluation of these algorithms using a prototype implementation that has been integrated in Infinispan. In opposition to the algorithms used in the current Infinispan implementation, based on Two-Phase Commit, our algorithms avoid deadlock scenarios. Our performance evaluation highlights speed-ups when using the proposed algorithms with respect to the native Infinispan replication mechanism. Keywords-Partial Replication, Distributed Memory, Transactional Memory, Atomic Multicast