FGSCM: A Fine-Grained Approach to Transactional Lock Elision

• Autor(a) principal: Sousa, Gustavo [UNESP]
• Data de Publicação: 2017
• Outros Autores: Baldassin, Alexandro [UNESP]
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1109/SBAC-PAD.2017.22; http://hdl.handle.net/11449/170591
• topic: Hardware transactional memory; Lock elision; Lock removal; Parallel programming; Transactional lock elision

Speculative Lock Elision (SLE) is a technique that allows critical sections to be executed optimistically by eliding the lock operation and enabling multiple threads to execute concurrently. In case of inconsistencies, the hardware automatically rolls back the execution and pessimistically acquires the original lock during runtime. The decision to elide the lock in SLE is performed transparently at the microarchitecture level and, although being convenient, it may sometimes hurt performance. To avoid that case, researchers have investigated Transactional Lock Elision (TLE), in which software-controlled hardware transactions are used instead, allowing the creation of policies and heuristics to manage lock elision. Typical implementations of TLE make use of a single lock to serialize the execution in case the original lock cannot be elided, which can potentially degrade performance. In order to improve on such cases, this paper proposes the Fine-Grained Software-assisted Conflict Management (FGSCM) scheme, a TLE technique that employs multiple locks so as to avoid unnecessary serialization of the code. The main idea of FGSCM is that not all threads that conflict inside a critical section are acessing the same region of shared memory. By automatically assigning distinct locks to these threads according to the memory section they access, the level of concurrency can be increased. In this paper we formalize FGSCM and provide an in-depth performance evaluation using a microbenchmark to stress several conflict behaviors. Our initial results with a prototype implementation using Intels Restricted Transactional Memory (RTM) are encouraging. With a quadcore machine, we observed an average performance gain of 11% compared to the single-auxiliary-lock SCM and 36% compared to a standard lock scheme, both for typical read-dominated workloads.