Public Thesis defense - ICTEAM

Speculation in Partially-Replicated Transactional Data Stores by Zhongmiao LI

Pour l’obtention du grade de Docteur en sciences de l’ingénieur et technologie

Distributed data storage systems are the key building blocks of today’s large-scale online services. While online services are increasingly employing geo-distribution and partial replication, state of the art approaches to enforce transactional consistency suffer from severe bottlenecks. This dissertation investigates the use of speculative techniques to enhance performance of partially-replicated, geo-distributed transactional data stores. With the term speculation, we refer to the possibility of exposing the updates of uncommitted transactions to other transactions and/or to external clients in order. We apply speculation techniques to two fundamental approaches to develop replicated transactional data stores, namely Deferred Update Replication (DUR) and State Machine Replication (SMR).

In DUR-based systems, transactions are firstly executed in a node and then propagated to other nodes for global verification, during which pre-commit locks are held on data updated by transactions. The global verification phase can severely throttle system throughput. We tackle this problem by introducing Speculative Transaction Replication (STR), a DUR protocol that exploits speculative reads to greatly reduce the ‘effective duration’ of pre-commit locks, thus enhancing the performance of partially-replicated, geo-distributed data stores. However, the indiscriminate use of speculative reads can expose applications to concurrency anomalies. Therefore, we introduce Speculative Snapshot Isolation (SPSI), an extension

of Snapshot Isolation (SI), which specifies desirable atomicity and isolation guarantees when using speculative processing techniques. Our experimental study shows that STR yields up to 11× throughput improvements over state-of-the art approaches.

In SMR-based systems, transactions first undergo an ordering phase, then replicas have to guarantee that the result of transaction execution is equivalent to a serial execution according to the produced order from the ordering phase. To ensure this guarantee, existing approaches use a single-thread to execute or serialize transactions, which severely limits throughput. We introduce Sparkle, a deterministic concurrency control designed for Partially-Replicated State Machines (PRSMs), which untaps the potential parallelism of modern multi-core systems through the use of speculative technique. The key contribution of SPARKLE is a set of techniques that can greatly minimize the frequency of misspeculations and the cost associated with correcting them. Our evaluation shows that SPARKLE achieves up to one order of magnitude throughput gains when compared to state of the art systems.

Jury members :

• Prof. Peter Van Roy (UCLouvain), promoteur
• Prof. Paolo Romano (IST, Portugal), promoteur
• Prof. Charles Pecheur (UCLouvain), président
• Prof. Etienne Riviere (UCLouvain), secrétaire
• Prof. Bruno Martins (IST, Portugal)
• Prof. Fernando Pedone (USI, Suisse)
• Prof. Danny Hugues (KUL, Belgique)

Pay attention :

The public defense of Zhongmiao Li scheduled for Friday 03 July at 04:00 p.m will indeed take place in the form of a video conference Teams.

Télécharger l'annonce