TCP congestion avoidance for selective flows in SDN

Abstract

We propose a system and a method for network-based flow-control to avoid Transmission Control Protocol (TCP) congestion for 'selective flows' in Software Defined Networks (SDN). In the proposed system, the controller, which has a global visibility of the network state uses switch data such as queue fullness and packet sojourn times, to enforce TCP flow control on certain flows, called 'ordinary flows,' unbeknown to end hosts. The objective is to prevent important or delay-sensitive so called selective flows to suffer from congestion collapse, which causes them to significantly reduce rate, and hence, the quality of experience for end-users. Our method achieves network-based TCP acknowledgment (ACK) pacing without modifying end-host TCP stack. To this effect, the controller sends new OpenFlow commands to request certain switches to modify some ACK header parameters and/or behavior of ACK packets of ordinary flows that share the same switch queues with the selective flows. Results of the experiments performed on a real testbed are provided to demonstrate the feasibility of the proposed method.