TRANSACTIONAL PROGRAMMABILITY AND PERFORMANCE BY

Transactional memory is a promising technique for multithreaded synchronization and con-currency which has attracted much interest in recent years. Compared to locks, transactions can offer increased concurrency, composability, and deadlock freedom; yet their speculative nature introduces problems not seen in lock-based code, and these have limited the adoption of transactions as a serious alternative to locks. Specifically, it is not clear how transactions can contain side-effecting actions like I/O, nor how they can scale to arbitrary footprints and durations while retaining acceptable performance and a consistent and desirable pro-gramming semantic. I believe that addressing problems like these is of first importance to programmers consid-ering using transactions. In this work, I seek to extend the programmability of transactions by exploring and addressing these problems. I begin by exploring the domain of hardware support for software transactional memory. I first show how a general-purpose user-mode memory protection system can be used to provide strong isolation to software transactional memory systems with little hardware and minimal performance overhead. Then I show ho