A practical and scalable congestion control scheme for high-performance multi-stage buffered switches

One of the challenging problems for multi-stage buffered switching is the performance degradation due to the saturation tree congestion inside the switch when traffic destined for some output ports exceeds their link capacity (i.e., hotspots) and blocks other traffic destined for non-overloaded output ports. In previous work [18], we have proposed HOPE, an effective congestion control scheme, in the 3-stage Clos Network on Chip (NOC). HOPE proactively regulates traffic destined for each output by estimating the number of their backlogged packets in the network and applying a simple stop-and-go mechanism to prevent hotspot traffic from jamming the internal links between the stages. The effectiveness of HOPE in NOC has motivated us to apply it in the multistage buffered switches. Different from an NOC, where Switch Modules (SMs) are all on the single chip, the SMs in a multi-stage buffered switch are separated from each other for a distance up to 100 m. This significantly increases the hardware complexity of HOPE. In this paper, we address the implementation challenges when applying HOPE in the 3-stage Clos network switch. In particular, we propose a scalable traffic measurement mechanism to approximate the backlogged traffic for each output port by taking advantage of the property of Clos network that traffic is evenly distributed among central SMs. We also design an efficient messaging system to notify input sources upon congestion status updates. Simulation results with different traffic patterns show that HOPE can isolate hotspot traffic from non-hotspot traffic, achieve max-min fairness among different traffic types, and provide low latency for non-hotspot traffic and high throughput for hotspot traffic.