Reducing cache invalidation overheads in wormhole routed DSMs using multidestination message passing

Abstract

Current generation distributed shared memory (DSM) systems use point-to-point (unicast) messages for cache invalidations. This incurs a large number of control messages, heavy network traffic, and high occupancy at home nodes. This paper introduces a new approach to reduce these overheads by using multidestination-based reservation and gather worms for distributing invalidation requests and collecting acknowledgments. Different grouping schemes to generate multidestination worms on networks supporting deterministic (e-cube) or adaptive (turn-model) routing are investigated to implement the fully-mapped cache-coherence protocol. For different applications on a 2D mesh system, our simulation results indicate that up to 15% reduction in overall execution time can be achieved by using multidestination messages.