Time domain performance evaluation of adaptive hybrid cache coherence protocols

Adaptive hybrid cache coherence protocols use both the write-invalidate mechanism and the write-update mechanism to maintain coherence among copies of data objects. Each of these protocols implements a decision function that chooses the appropriate mechanism in order to improve their performance. In most existing solutions, decision functions are based on communication traffic. Moreover, the authors of the adaptive protocols use communication traffic as a performance measure in their papers. In contrast, in this paper we present the results of a performance evaluation of adaptive hybrid cache coherence protocols in both the traffic domain and in the time domain. We compare three adaptive protocols with pure write invalidate and pure write-update protocols. Let r_WI, r_WU, and r_A be the average communication traffic per access for the write-invalidate protocol, the write-update protocol, and the adaptive protocol, respectively. The adaptive protocol minimizes the traffic if r _A≈min(r_WI, r_WU). Similarly, the adaptive protocol minimizes the access latency if t_A≈min(t _WI, t_WU), where t_WI, t_WU, and t_A are the access latencies for the write-invalidate protocol, the write-update protocol, and adaptive protocol, respectively. For some of the workload parameters the adaptive protocols minimize both traffic and access latency. However, we also present and analyze the workload parameters for which adaptive protocols minimize communication traffic, but fail to minimize the access latency