Updating Data-Center Network With Ultra-Low Latency Data Plane

摘要

Due to the asynchronous and distributed nature of the data plane, the transition from the initial to final state may result in serious transient congestion in software-defined networking (SDN). Moreover, with the rapid development of ultra-low latency network in data-centers, the control loop between the control and data plane becomes much longer than the ultra-low latency communication in the data plane. In this case, traffic surges significantly in the data plane during the network reconfiguration process, and it becomes harder for the SDN controllers to manipulate the update operations as they expect. In this paper, we consider the traffic variation during the network update and model it as a novel minimum demand violation problem (MDVP). Later, we prove its hardness, and propose a heuristic approximation algorithm to approach the optimal result. Our method brings flexibility for network operators to make a trade-off between the network congestion and update speed. Experiments show that our method can halve the intermediate network states and reduce the demand violation ratio by 36.7 % compared to the state-of-the-art.