On-chip networks for manycore architecture

Abstract

Over the past decade, increasing the number of cores on a single processor has successfully enabled continued improvements of computer performance. Further scaling these designs to tens and hundreds of cores, however, still presents a number of hard problems, such as scalability, power efficiency and effective programming models. A key component of manycore systems is the on-chip network, which faces increasing efficiency demands as the number of cores grows. In this thesis, we present three techniques for improving the efficiency of on-chip interconnects. First, we present PROM (Path-based, Randomized, Oblivious, and Minimal routing) and BAN (Bandwidth Adaptive Networks), techniques that offer efficient intercore communication for bandwith-constrained networks. Next, we present ENC (Exclusive Native Context), the first deadlock-free, fine-grained thread migration protocol developed for on-chip networks. ENC demonstrates that a simple and elegant technique in the on-chip network can provide critical functional support for higher-level application and system layers. Finally, we provide a realistic context by sharing our hands-on experience in the physical implementation of the on-chip network for the Execution Migration Machine, an ENC-based 110-core processor fabricated in 45nm ASIC technology.