A low-latency scalable locking algorithm for shared memory multiprocessors

In this paper we deal exclusively with spin locks for small- to medium-scale shared memory multiprocessors. The performance of a lock depends on the contention for the critical section being protected by the lock. The selection of the most appropriate lock for a critical section is difficult because the level of contention can vary dynamically. We verify that no current lock has the best performance for all critical sections. We present a new lock, called the shared array lock, which alleviates this problem. The shared array lock belongs to the class of locks based on queuing of processors waiting for the lock. Unlike existing queue-based locks which suffer from high latency in low contention situations, experimental results show that the shared array lock has low latency and good scalability