Congestion-oriented shortest multipath routing

We present a framework for the modeling of multipath routing in connectionless networks that dynamically adapt to network congestion. The basic routing protocol uses a short-term metric based on hop-by-hop credits to reduce congestion over a given link, and a long-term metric based on end-to-end path delay to reduce delays from a source to a given destination. A worst-case bound on the end-to-end path delay is derived under three architectural assumptions: each router adopts weighted fair queueing (or packetized generalized processor sharing) service discipline on a per destination basis, a permit-bucket filter is used at each router to regulate traffic flow on a per destination basis, and all paths are loop free. The shortest multipath routing protocol regulates the parameters of the destination-oriented permit buckets and guarantees that all portions of a multipath are loop free.