Abstract
Rapid advances in interconnection networks in multiprocessors are closing the gap between computation and communication. Given this trend, how can we utilize fast interconnects? This study proposes an enhanced CC-NUMA architecture, called Depot-NUMA, which views the congregation of the private caches in all nodes as a large remote access cache. Fast interconnects allow a missing block to be fetched from the private caches of other sharing nodes rather than from the home node. Issues involved in designing Depot-NUMA are also discussed, and a novel routing scheme, called multi-hop, is proposed to communicate between the potential sharers and fetch a missing block from their private caches. The sharers are specified based on a stride function to exploit network locality in the system. The proposed Depot-NUMA design requires only modest modification to the node controller and coherence protocol. Additionally, the interconnect fabric can be constructed using existing and unmodified commodity interconnects. Furthermore, the application-driven study reveals that Depot-Numa can reduce the read stall time by up to 41% and is competitive compared to a CC-NUMA with a large local cache.