A Time Variant Fluid Model for DCQCN Congestion Control Protocol

The development of datacenter applications leads to the need for high throughput and ultra-low latency from the network. Remote Direct Memory Access (RDMA), which is becoming prevalent in datacenter networks in recent years, is capable to provide ultra-low latency and high throughput with less CPU utilization. Congestion control is of great importance for RDMA for its lossless feature. Datacenter QCN (DCQCN) is the most commonly used congestion control mechanism and is deployed in large-scale datacenter networks. To ensure the stability and performance of DCQCN, a fluid model is essential. However, due to neglecting time variant factors in the queue dynamic, the existing fluid model is not precise enough. In this paper, we propose a time variant fluid model which can be more effective to evaluate the stability and performance of DCQCN. In this model, the variation of the queue delay and system cut-off frequency is considered. As a result, the non-monotonic stability behavior is eliminated and a more reasonable and accurate evaluation of the congestion control system is acquired.