PFT: A Congestion Avoidance Method based on Proactive Flow Throttling at Endpoints

Network congestion will result in severe performance degradation in a large-scale lossless interconnection network if no effective countermeasure is taken. An implementable, effective and cost-efficient design is crucial to practical network congestion control. Most existing reactive and proactive congestion control schemes still suffer from complicated congestion signals and parameter tuning, which significantly increases implementation complexity and the risk of incorrect settings that cause poor performance. In this work, we make a successful attempt to present a novel congestion control that minimizes the complexity and cost of protocol implementation. We explore proactive flow throttling (PFT) techniques at host network interface. PFT merely uses the local credit information at the endpoint node. Our design can quickly sense the network congestion and respond to it by proactive flow throttling with simple parameter configuration. This method has be successfully implemented and verified in one high-performance network interface chip at an extremely low cost. We use the state-of-the-art benchmark dedicated to network congestion testing to evaluate our design. The extensive results show that our method can quickly detect the occurrence of network congestion, and effectively perform congestion control in a real-time manner, significantly improving the application performance.