Adaptively weighted round-robin arbitration for equality of service in a many-core network-on-chip

This study presents a technique called adaptively weighted round-robin (RR) arbitration for equality of service in a many-core network-on-chip. The authors concentrate on the network congested with various traffic patterns generated by the applications running on the system. It exploits the deterministic properties of the interconnection network - the topology and the routing algorithm - to achieve the global fairness in terms of service provided to each node with less resource requirements compared with previous work. The weights for input arbitration can be adjusted to make the network better adapted to various traffic patterns. It requires no additional information in packet headers. The hardware overhead is minimal, requiring only several small counters in addition to a typical RR arbiter. The critical path delay is also reduced due to its simplicity. The authors show the effectiveness by implementing RTL models of the routers and synthesizing them with 32/28 nm process technology. SPEC CPU2006 benchmark applications are executed in multi-programmed manner to show that the approach results in outstanding equality-of-service characteristics for real applications.