Improved Single Global Lock Fallback for Best-effort Hardware Transactional Memory

Intel’s Haswell and IBM’s Blue Gene/Q and System Z are the first commercially available systems to include hardware transactional memory (HTM). However, they are all best-effort, meaning that every hardware transaction must have an alternative software fallback path that guarantees forward progress. The simplest and most widely used software fallback is a single global lock (SGL), in which aborted hardware transactions acquire the SGL before they are re-executed in software. Other hardware transactions need to subscribe to this lock and abort as soon as it is acquired. This approach, however, causes many hardware transactions to abort unnecessarily, determining even more transactions to fail and resort to the SGL. In this paper we suggest improvements to the simple SGL fallback. First, we use lazy subscription to reduce the rate of SGL acquisitions. Next, we propose fine-grained conflict detection mechanisms between hardware transactions and a software SGL transaction. Finally, we describe how our findings can be used to improve future generations of HTMs.