Accelerating Transactional Memory by Exploiting Platform Specificity

Transactional Memory (TM) is one of the most promising alternatives to lock-based concurrency, but there still remain obstacles that keep TM from being utilized in the real world. Performance, in terms of high scalability and low latency, is always one of the most important keys to general purpose usage. While most of the research in this area focuses on improving a specific single TM implementation and some default platform (a certain operating system, compiler and/or processor), little has been conducted on improving performance more generally, and across platforms.We found that by utilizing platform specificity, we could gain tremendous performance improvement and avoid unnecessary costs due to false assumptions of platform properties, on not only a single TM implementation, but many. In this dissertation, we will present our findings in four sections: 1) we discover and quantify hidden costs from inappropriate compiler instrumentations, and provide sug- gestions and solutions; 2) we boost a set of mainstream timestamp-based TM implementations with the x86-specific hardware cycle counter; 3) we explore compiler opportunities to reduce the transaction abort rate, by reordering read-modify-write operations — the whole technique can be applied to all TM implementations, and could be more effective with some help from compilers; and 4) we coordinate the state-of-the-art Intel Haswell TSX hardware TM with a software TM "Cohorts", and develop a safe and flexible Hybrid TM, "HyCo", to be our final performance boost in this dissertation.The impact of our research extends beyond Transactional Memory, to broad areas of concurrent programming. Some of our solutions and discussions, such as the synchronization between accesses of the hardware cycle counter and memory loads and stores, can be utilized to boost concurrent data structures and many timestamp-based systems and applications. Others, such as discussions of compiler instrumentation costs and reordering opportunities, provide additional insights to compiler designers. Our findings show that platform specificity must be taken into consideration to achieve peak performance.