Improving coherence protocol reactiveness by trading bandwidth for latency

Abstract

This paper describes how on-chip network particularities could be used to improve coherence protocol responsiveness. In order to achieve this, a new coherence protocol, named LOCKE, is proposed. LOCKE successfully exploits large on-chip bandwidth availability to improve cache-coherent chip multiprocessor performance and energy efficiency. Provided that the interconnection network is designed to support multicast traffic and the protocol maximizes the potential advantages that direct coherence brings, we demonstrate that a multicast-based coherence protocol could reduce energy requirements in the CMP memory hierarchy. The key idea presented is to establish a suitable level of on-chip network throughput to accelerate synchronization by two means: avoiding the protocol serialization, inherent to directory-based coherence protocol, and reducing average access time more than in other snoop-based coherence protocols, when shared data is truly contended. LOCKE is developed on top of a Token coherence performance substrate, with a new set of simple proactive policies that speeds up data synchronization and eliminates the passive token starvation avoidance mechanism. Using a full-system simulator that faithfully models on-chip interconnection, aggressive core architecture and precise memory hierarchy details, while running a broad spectrum of workloads, our proposal can improve both directory-based and token-based coherence protocols both in terms of energy and performance, at least in systems with up to 16 aggressive out-of-order processors in the chip.