In this paper, we develop algorithms for decentralized traffic engineering when several classes of services are to be provided. More precisely, given an utility function of a very general form, we provide a large class of distributed control laws which converge to the maximum of the utility function subject to the network constraints and service requirements. Each ingress node adapts its sending rates and distributes the traffic among the multiple available paths independently and with minimum feedback from the network. These laws allow not only for best effort traffic (only type of traffic in today’s Internet) but also assured service traffic where a given target rate has to be achieved in average.