UNIVERSIDADE DE LISBOA INSTITUTO SUPERIOR TÉCNICO Autonomic Replicated Software Transactional Memory

Software Transactional Memory (STM) systems have emerged as a powerful paradigm to develop concurrent applications. By sparing the programmer from the burden of explicitly dealing with low-level concurrency mechanisms, STMs increase the code reliability and shorten the development time. For scalability and fault-tolerance reasons, it is relevant to build distributed implementations of this paradigm allowing the deployment of transactional systems in environments such as the cloud. Replication clearly plays a role of paramount importance in these platforms, as it represents the key mechanism to ensure data durability in face of unavoidable node failures. However, there is a myriad of techniques to maintain the consistency of the replicated data, each performing differently according to the workload and there is no “one-size fits all” technique that achieves optimal performance in any scenario. In addition, due to the growing complexity of these systems and unpredictability of the workloads, manual management becomes a complex, tiresome and error-prone task. To circumvent these issues, this thesis investigates an autonomic architecture for the replication of STMs that supports the change of its replica consistency protocol according to the workload, in order to provide the best throughput possible, using techniques that do not require human intervention. First, it studies the use of machine learning to support the autonomic management of the replica consistency protocols. Then, the thesis presents a solution that relies on two machine learning approaches to drive the adaptation in a distributed STM that is able to switch between certification-based protocols. Finally, it proposes a framework for autonomic distributed STMs that supports multiple replica consistency protocols and two switching mechanisms: one forcing the system to stop processing transactions while the protocol switch is in progress, and another which allows the programmer to define how the two switching protocols can co-exist during the transition period. Palavras Chave Keywords