Evaluating and Optimizing Real-Time Software Transactional Memory

Software transactional memory (STM) is a proposed solution to the challenge of developing correct concurrent code. STM allows programmers to annotate sections of their code that need to be synchronized, and the STM implementation resolves synchronization issues behind the scenes. One application domain where STM can be particularly useful is real-time systems, where schedulability is a crucial metric is system certification. However, STM usually relies on retries to resolve contention, which makes theoretical worst-case behavior, and thus schedulability, highly pessimistic and therefore impractical in the real-time application domain. Previous work on real-time STM has failed to give both theoretical and practical solutions to this problem. Work in this thesis is part of a large effort to present real-time STM that is both schedulable and high-performing, achieved using a retry-free, and thus entirely lock-based, STM implementation. This work details multiple metrics for evaluating retry-free STM compared to a known retry-based lock-based solution, as well as presents optimizations to a locking protocol for increased performance in this context. Evaluations show a retry-free STM implementation with optimized locking protocols is significantly more schedulable and higher performing on single-socket machines than other lock-based STM.Bachelor of Scienc