Adaptive Cluster Throttling: Improving High-Load Performance in Bufferless On-Chip Networks

Higher core counts and increasing focus on energy efficiency in modern Chip Multiprocessors (CMP) have led to renewed interest in simple and energy-efficient Network-on-Chip (NoC) designs. Several recent proposed designs trade off network capacity for efficiency, based on the observation that traditional networks are overprovisioned for many workloads. Bufferless routing is one such example. However, when the application workload requires high interconnect performance, the inefficiencies of bufferless interconnects can cause significant performance degradations. Previous work has tackled various issues with bufferless routing, but little work has been done to improve performance at high network load. Fundamental improvements in bufferless network performance at high load could extend the benefits of lower energy and smaller die area to a wider range of potential applications. In this work, we propose ACT (Adaptive Cluster Throttling), a source-throttling mechanism that provides better system performance and fairness than the best current mechanisms on bufferless networks. By batching applications into clusters, and alternately throttling different clusters, ACT provides a chance for all applications to inject traffic into the network while maintaining control over total network load. We show 11.9% (10.2%) system performance gain on average with 14.5% (15.1%) improvement in fairness over 60 network-intensive workloads on a 4x4 (8x8) bufferless NoC. At high network load, ACT achieves nearly half the performance gain over a bufferless baseline that a conventional buffered network achieves, while reducing network power by 15.4% (5.4%).