Improving the Practicality of Transactional Memory a Dissertation Submitted to the Department of Electrical Engineering and the Committee on Graduate Studies of Stanford University in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Transactional Memory (TM) simplifies parallel programming by transferring responsibility for concurrency management from the programmer to the system. Nevertheless, a number of challenges must be addressed for TM to be widely adopted. This dissertation presents work towards improving the practicality of TM across three dimensions. The first challenge addressed is that of integrating TM with popular models for high-level parallel programming. The need for high-level programming environments for TM is ever increasing to improve the programmability of TM. To address this challenge, we propose OpenTM, a programming model that integrates TM into OpenMP, a widely-used parallel programming environment. The high-level language constructs of OpenTM simplify TM programming by abstracting low-level programming issues such as manual instrumentation of TM barriers and by allowing high-level management of TM programming options. OpenTM aims to deliver good performance with simple and portable high-level code. The second challenge addressed is that of supporting nested parallelism within transactions. To achieve the best possible performance on large-scale parallel systems, it is important to fully exploit the parallelism available at all levels. To address this challenge, we propose two TM systems with support for concurrent nesting: (1) NesTM, a software TM (STM) system that supports nested parallel transactions solely in software and (2) Filter-accelerated Nested TM (FaNTM), a hybrid TM system with lightweight hardware support for nested parallel transactions. NesTM presents a good model for concurrent nesting, and FaNTM provides a fast and practical implementation. In particular, the hardware filters of FaNTM provide continuous