Congestion and Flow Control in the Context of the Message-Oriented Protocol SCTP

Abstract

Congestion and flow control are key mechanisms used to regulate the load in modern packet networks. The new IETF Stream Control Transmission Protocol (SCTP) inherited these algorithms from the Transmission Control Protocol (TCP). Although the principles used are the same, some issues arise from the fact that SCTP operates message-oriented whereas TCP operates byte-stream oriented. SCTP also supports bundling of multiple small user messages into one SCTP packet. As a consequence, the overall overhead of an SCTP packet depends on the user message size and the number of user messages that are bundled into the packet. RFC 4960 defining SCTP does not specify whether the message specific headers have to be considered when updating the parameters for congestion control. We will show that neglecting the additional headers when calculating outstanding bytes can lead to unfairness towards TCP connections. We will also show that incorrect handling of the additional memory needed to process each message in the flow control calculations will lead to an exhaustion of the receiver window resulting in a huge amount of unnecessary retransmissions. Based on experiments with the flow control of the SCTP implementations available in several operating systems, we will identify the issues and analyze them by using simulations. As a result, we will present solutions that will lead to fairness towards TCP and reduce the number of retransmissions substantially. Although we will focus on SCTP, the results are also true for other message-oriented protocols using bundling.