Brief announcement: locally-accessible implementations for distributed shared memory multiprocessors

Abstract

We consider asynchronous multiprocessors that support the distributed shared memory (DSM) model. Algorithms for such multiprocessors exploit the ability to co-locate shared objects with particular processes in order to reduce the cost of accessing shared memory. When a shared object fits inside a single memory word and operations on it are supported directly through machine instructions, it can be made local to any process simply by fixing its physical address. We show that even if the shared object is not supported in hardware directly, it can always be simulated using a software implementation that behaves as though it is local to some designated process. That is, operations applied by the designated process on the implemented object access only local base objects, which is non-trivial when processes synchronize by busy-waiting. We also discuss time complexity bounds for such locally-accessible implementations.