Model Predictive Congestion Control for Data Center Networks

This paper is concerned with congestion control (CC) of data center networks (DCN) from a control theoretical perspective. We first establish a discrete-time state-space model to capture the behaviors between the queue length and congestion window (CWND) size. Leveraging this model, we propose a model predictive congestion control (MPCC) algorithm in remote direct memory access-enabled DCNs. As a window-based CC, MPCC dynamically adjusts the flow's CWND size according to the observed network conditions. To further enhance MPCC performance, an explicit MPCC (EMPCC) is also constructed, which can be pre-solved offline to reduce online computational cost guided by the explicit model predictive control (MPC) method. Through experiments ranging from bursty incast to real-world traffic patterns, we demonstrate that (E)MPCC significantly reduces occupied switch queue length and flow completion time while maintaining flow fairness and throughput, outperforming other well-known and widely adopted DCN-special CC algorithms such as DCTCP, DCQCN, TIMELY, SWIFT, and HPCC.