Exploiting multi-band transmission line interconnects to improve the efficiency of cache coherence in multiprocessor system-on-chip

Abstract

Main-stream general-purpose microprocessors integrate a growing number of cores on-chip, requiring high-performance interconnects and efficient cache coherence for data transmission and sharing. Conventional directory-based cache coherence has high indirection overhead, which adds to the critical path of data requests and lowers overall system performance. In fact, with globally shared high-performance interconnects, cache coherence could be optimized and the indirection overhead could be relieved. This paper explores the use of multi-band transmission lines to implement globally shared interconnects. Taking advantage of the aggregate frequency band resources, the proposed interconnect supports efficient multi-cast, and helps improve the efficiency of cache coherence with augmented parallelism. Coherence indirections are avoided, and we have seen an average of 17% boost in application performance, as well as an average of 18% throughput improvement compared to an implementation of conventional cache coherence on single-band transmission line based interconnects.