Modeling and Resolving Lock Contention for Multi-threaded Systems

Locks are efficient concurrent control mechanisms to ensure that shared resources are accessed by only a single thread in multi-threaded applications. However, for multi-core systems, lock usage leads to lock contention, which degrades program performance, increases system response time, and negatively affects scalability. This paper initially models lock contention from two aspects: static structure and dynamic behavior. Static constraints of lock contention that describe the contention among threads and the dynamic constraints that define the lock request behavior during contended process are presented. In addition, a contention-aware lock (CA-Lock) is proposed. The CA-Lock is an ordinary lock when there is no contention; it improves the concurrency by transforming into software transactional memory (STM) when the contention is serious. For those situations that are not fitted with STM, the CA-Lock can transform back into an ordinary lock. Primary experimental results show that the CA-Lock can track the best world of lock or STM.