ABDTR: Approximation-Based Dynamic Traffic Regulation for Networks-on-Chip Systems

Traffic regulation is an essential technology of networks-on-chip (NoC) to achieve communication performance guarantees with effective use of the system interconnect and low traffic delay. This paper presents approximation-based dynamic traffic regulation (ABDTR), which approximates part of traffic data instead of network transmission for mitigating network congestion based on the inherent error resilience of some applications. ABDTR drops a fraction of safe-to-approximate packet data that may aggravate network congestion before they are injected into network, and predicts the lost data in packet after being received in destination node. Traffic drop rate is fully adaptive to the traffic and network states. The regulation range of drop rate is a knob to control the trade-off between performance/energy efficiency and output quality. The proposed method is effective and can be implemented in hardware with small area. Experimental results have confirmed that the ABDTR can help reduce up to 44.4 percent average network delay with less than 10% loss in quality. The results also show ABDTR yields significant application speedup and NoC energy reduction for a wide range of quality-loss levels.