Locking Protocols for Large - ScaleCache - Coherent Shared Memory

Signiicant performance advantages can be realized by implementing a database system on a cache-coherent shared memory multiprocessor. An eecient implementation of a lock manager is a prerequisite for eecient transaction processing in multiprocessor database systems. To this end, we examine two approaches to the implementation of locking in a cache-coherent shared memory multiprocessor database system. The rst approach, shared-memory locking (SML), allows each node (processor) of the multiprocessor to acquire and release locks directly via the use of cache-coherent shared memory. The second approach, message-passing locking (MPL), typically requires messages to be sent to a lock manager, located on a remote node. Our empirical evaluation of these approaches on the KSR-1 multiprocessor indicates that for most database locking traac patterns, the performance of SML is substantially superior to that of MPL. For instance, when contention is high, the performance of SML is nearly an order of magnitude better than MPL.