Efficient Protocols for Replicated Transactional Systems

Sebastiano Peluso

Abstract:
Over the last years several relevant technological trends have significantly increased the relative impact that the inter-replica synchronization costs have on the performance of transactional systems. Indeed, the emergence of technologies like Transactional Memory, Solid-State Drives and Cloud computing has exacerbated the ratio between the latencies of replication coordination and transaction processing. The requirements of these environments harshly challenge state of the art techniques for replication of transactional systems, raising the need for rethinking existing approaches to this problem.
This dissertation advances the state of the art on replicated transactional systems by presenting a set of innovative replication protocols designed to achieve high efficiency even in such challenging scenarios.
More in detail, four transactional replication protocols are proposed, which tackle the aforementioned issues from various angles. The first two cope with full replication scenarios, and exploit orthogonal techniques, such as speculation and transaction migration, which allow for amortizing, in different ways, the impact of distributed coordination on system performance. The other two proposals explicitly cope with the issue of scalability, by introducing the first genuine partial replication techniques that support abort-free read-only transactions while ensuring, respectively, One-Copy Serializability and Extended Update Serializability. The core of these protocols is a distributed multi-version concurrency control algorithm, which relies on a novel logical clock synchronization mechanism to track, in a totally decentralized (and consequently scalable) way, both data and causal dependency relations among transactions. The trade-offs arising across the different presented solutions are also discussed and experimentally evaluated by integrating them into state of the art academic and industrial transactional platforms.

BibTeX Entry: