Power: Multi-Capatibility Adaptive Routing for Network-an-Chips
Xiaoyun Zhang · 2023
Effective adaptive routing is essential for optimal network-on-chips (NoCs) performance, especially for complex and diverse traffic patterns of emerging applications on many-core architectures. However, adaptive routing still encounters significant challenges. It is well known that adaptive routing should have maximizing routing adaptiveness, but it breaks the uniform traffic pattern, reduces the network performance, and can be endpoint congestion in the presence of non-uniform traffic patterns (such as hotspot traffic patterns). We found that most existing adaptive routing fails to be compatible with multiple traffic patterns and network congestion at the same time. To address this limitation, we propose the Power _ a new adaptive routing algorithm that has a strengthen routing adaptiveness and backpressure capability. The Power algorithm can effectively isolate endpoint congestion by avoiding the path of packets that are about to reach the destination to reduce congestion spreading and can achieve dynamic load-balancing by regulating routing adaptiveness when meeting heavy uniform traffic. We conduct extensive experiments against Footprint routing algorithms to evaluate our routing algorithm, and our simulation results show that our design can be improved by up to 36.15% under heavy synthetic traffic patterns.