Adaptive cache coherence over a high bandwidth broadband mesh network

Abstract

Networks have traditionally been an obstacle to high performance distributed computing. Specific problems are insufficient bandwidth and long transaction latencies. While pipelining data can achieve high bandwidth, it does nothing for latency which is still a bottleneck in performance. One approach is to develop a cache coherence protocol which exploits recurring data sharing patterns to reduce the impact of latency. This paper proposes an adaptive cache coherence protocol which detects producer–consumer type sharing and maintains coherence on only those cache blocks which exhibit producer–consumer sharing via updates rather than invalidates. Execution driven simulations of this protocol show improved performance compared to a standard write-invalidate protocol protocol and a competitive update protocol. When there are no access patterns to exploit, the protocol does not degrade performance. When there is producer–consumer type sharing, the proposed protocol runs benchmarks up to 30% faster than the better of either write-invalidate or competitive update. As a side-effect, it shows improved tolerance of increasing network latency.