Consensus Based Value Caching Network-on-Chip (CVC-NoC))

Abstract

Caches in on-chip systems are conventionally built around the two principles: spatial and temporal locality, which respectively rely on the likelihood of referencing the same address (or its neighbours) in the near future. Value locality was coined as a third facet of locality in 1996. This type of locality indicates the likelihood of the recurrence of the same value in a storage/communication system. With the increase in demand for chip-level multiprocessor (CMP) systems, in almost every application, the performance of Network-on-chips (NoCs) plays a critical role. In this paper, we propose a value locality-based mechanism for value compression in an NoC. We have developed and tested a consensus-based approach to design a scalable and resource-friendly table-based compression of frequent values. Our experiments show that for an 8x8 mesh-based NoC, our value locality scheme achieves a reduction of around 46% in power and 8% in packet latency using single-threaded workloads, compared to a baseline NoC without value caching. For multithreaded workloads, we achieved a reduction of around 34% in NoC power and 6% in packet latency. For the best cache configuration, these latency numbers are at least 2% better than a value caching scheme that only compresses the zero value.