Abstract
With the advent of multicore systems, the transactional memory (TM) concept has attracted much interest from both academy and industry. Indeed, by supporting the ACI (Atomicity, Consistency and Isolation) proprieties, the TM concept eases parallel programming and avoids the severe problems of lock-based methods such as deadlock situations and priority inversion. In addition, unlike lock-based methods, TM allows several transactions to access resources in parallel, and thus increases the system's bandwidth. Moreover, TM embeds a transaction's scheduler which either rollbacks the transaction when the conflict is detected, or commits the transaction on success. The thesis's objective is to study the TM's adaptation to soft real-time systems in which processes must complete within deadlines as far as it is possible. Up to now, the scheduling of real-time transactions within TM has not been studied. To address this issue, we first make an experimental and comparative study to show whether the TM is suitable for real-time multicore systems. In particular, we evaluate the transaction's execution time variation when accessing shared resources. Second, we propose a novel solution which introduces a real-time transactional model within TM and we design and implement a real-time software transactional memory named RT-STM. This one integrates new synchronization protocols which allows to prioritize the shared resources' accesses according to the processes' urgency. Finally, we show how to make RT-STM suitable for firm real-time systems, proposing some adaptations allowing to guarantee a certain level of quality of service (QoS) to processes sharing resources.