View caching: Efficient software shared memory for dynamic computations

Software distributed shared memory (DSM) techniques, while effective on applications with coarse-grained sharing, yield poor performance for the fine-grained sharing encountered in applications increasingly relying on sophisticated adaptive and hierarchical algorithms. Such applications exhibit irregular communication patterns unsynchronized with computation, incurring large overheads for synchronous (request-reply) DSM protocols that require responsive processing of coherence messages. We describe a new DSM framework, View Caching, that addresses this problem by utilizing application knowledge of data access semantics to enable the construction of low-overhead, asynchronous coherence protocols. Experiments on the Gray T3D show that view caching enables efficient execution of fine-grained irregular applications, reducing both coherence overheads and idle time to improve performance by up to 35% over a weakly-consistent DSM implementation.