Fast Timescale Traffic Engineering in MPLS Networks

Abstract

Traffic engineering can be used to solve congestion through efficient traffic distribution. The network with differentiated classes of service is the main focus in this work. In general, providers offer QoS guaranteed in high priority class by serving the packets from that class before the packets from low-priority class. This strategy will work as long as the high priority traffic share is low (5-10%). We are interested in increasing that high priority traffic share. However, the large share of high priority class could lead to a QoS-violation risk. Moreover, it could decrease the goodput of the low-priority traffic as a result of its elastic nature, for example, from the congestion control in TCP. We present two works based on the fast timescale traffic engineering. The first work focuses on the flow migration among the parallel working paths. The potential usefulness of fast timescale migration control is explored with packet-level simulation and experimental testbed. The migration can improve the QoS of real-time traffic by shifting some of the flows out of the congested links. In the design of the migration controller, we will also consider the interaction between the controller and the TCP congestion control mechanism. We formulate the partially observed Markov Decision Process (MDP) problem and solve it using the dynamic programming technique. To cope with the well-known curse of dimensionality, we also present sub-optimal controllers, which scale well with our large network. Our second work tried to exploit the unused backup path for QoS improvement. Many network providers already setup a backup path for each working path. This backup path will protect the working path upon its failure. Without any failures, the providers fill the backup path with lower-priority traffic to increase their network's throughput. In general, the backup paths will not be used for high priority traffic in normal condition even with the congestion situation of the high priority traffic. We propose the scheme to duplicate the high priority packets to backup path in the congestion condition. With a small throughput degradation of low-priority traffic, we could gain significant QoS improvement of high priority class.