Eecient, Strongly Consistent Implementations of Shared Memory

We present linearizable implementations for two distributed organizations of multiprocessor shared memory. For the full caching organization , where each process keeps a local copy of the whole memory, we present a linearizable implementations of read/write memory objects that achieves essentially optimal eeciency and allows quantitative degradation of the less frequently employed operation. For the single ownership organization, where each memory object is \owned" by a single process which is most likely to access it frequently, our linearizable implementation allows local operations to be performed much faster (almost instantaneously) than remote ones. We suggest to combine these organizations in a \hybrid" memory structure that allows processes to access local and remote information in a transparent manner, while at a lower level of the memory consistency system, diierent portions of the memory are allocated to employ the suitable implementation based on their typical usage and sharing pattern .