Extending TM Primitives using Low Level Semantics

Abstract

Transactional Memory (TM) has recently emerged as an optimistic concurrency control technique that isolates concurrent executions at the level of memory reads and writes, therefore providing an easy programming interface. However, such transparency could be overly conservative from an application-level perspective. In this work, we propose an extension to the classical TM primitives (read and write) to capture program code semantics (e.g., conditional expressions) while maintaining the same level of programming abstraction. We deployed this extension on two state-of-the-art STM algorithms and integrated it into the GCC compiler and the RSTM software framework. Results showed speedups of up to 4x (average 1.6x) on different applications including micro benchmarks and STAMP.