Lock improvement technique for release consistency in distributed shared memory systems

Distributed shared memory allows processes to view the physically distributed memory as a globally shared virtual memory. Lazy release consistency (LRC), among known techniques, is an efficient software model proposed for distributed shared memory. It relies heavily on lock synchronization to maintain data coherency. The lock scheme used in LRC, however, conducts many interrupt invocations on the remote processors, which in turn steal effective cpu cycles from remote processors, thus prolonging the lock acquisition time and the total elapsed time of application programs. In this paper, a lock improvement technique is proposed to alleviate interrupt invocations caused by the lock acquire operations, leading to reduction in the lock acquisition time and the overall program execution time. Our improvement technique was evaluated under the TreadMarks' framework using four applications, where TreadMarks is a distributed shared memory system based on LRC. The experimental results indicate that our technique improves the lock acquisition time over TreadMarks on a network of workstations by more than 14% for one application.