Deploying Hardware Locks to Improve Performance and Energy Efficiency of Hardware Transactional Memory

Abstract

In the search for new paradigms to simplify multithreaded programming, Transactional Memory (TM) is currently being advocated as a promising alternative to lock-based synchronization. Among the two most important alternatives proposed for conflict detection and data versioning in today’s Hardware Transactional Memory systems (HTMs), the Lazy-Lazy one allows increased concurrency, potentially bringing higher performance levels in most cases. Unfortunately, the implementation of the commit protocol in Lazy-Lazy systems results in increased complexity and has severe impact on performance and energy consumption. In this work, we propose GCommit, an efficient and low cost hardware implementation of the SEQ commit protocol based on the use of hardware locks. Specifically, GCommit deploys hardware locks to ensure exclusive access to shared data at commit time. Implementing this functionality using dedicated hardware brings important benefits in terms of execution time as well as energy consumption with respect to traditional commit protocols that use the general-purpose interconnection network . Additionally, our proposal has negligible requirements in terms of area. Results for a 16-core CMP show that the GCommit protocol obtains average reductions of 15.7% and 13.7% in terms of execution time and energy consumption, respectively, compared with a traditional implementation of Scalable TCC with SEQ, a high-performance commit protocol proposed in the literature.