Partial replication in distributed software transactional memory

Abstract

Distributed software transactional memory (DSTM) is emerging as an interesting alternative for distributed concurrency control. Usually, DSTM systems resort to data distribution and full replication techniques in order to provide scalability and fault tolerance. Nevertheless, distribution does not provide support for fault tolerance and full replication limits the system’s total storage capacity. In this context, partial data replication rises as an intermediate solution that combines the best of the previous two trying to mitigate their disadvantages. This strategy has been explored by the distributed databases research field, but has been little addressed in the context of transactional memory and, to the best of our knowledge, it has never before been incorporated into a DSTM system for a general-purpose programming language. Thus, we defend the claim that it is possible to combine both full and partial data replication in such systems. Accordingly, we developed a prototype of a DSTM system combining full and partial data replication for Java programs. We built from an existent DSTM framework and extended it with support for partial data replication. With the proposed framework, we implemented a partially replicated DSTM. We evaluated the proposed system using known benchmarks, and the evaluation showcases the existence of scenarios where partial data replication can be advantageous, e.g., in scenarios with small amounts of transactions modifying fully replicated data. The results of this thesis show that we were able to sustain our claim by implementing a prototype that effectively combines full and partial data replication in a DSTM system. The modularity of the presented framework allows the easy implementation of its various components, and it provides a non-intrusive interface to applications.