Path Computation for Incoming Interface Multipath Routing

Currently used IP routing protocols calculate and only use a single path between two nodes of a network, or in the best case, only paths with the same cost (with OSPF2 or IS-IS extension: ECMP). If we want to use the underlying physical network with multipath routing efficiently, the loopfreeness of the used paths has to be guaranteed especially with distributed computation. Indeed different types of traffic engineering with source computation like OMP-MPLS or MATE-MPLS do not have to pay attention to loops. However the positioned paths, with CR-LDP or RSVP-TE for example, are not flexible enough to support strong load oscillations. Load balancing is only possible on the ingress node which labels the different paths (with a hash function to avoid packet mis-ordering for TCP traffic) even though the congestion spot can be very far from the ingress node. This is why distributed techniques can react more quickly to prevent congestion when possible. But such techniques do not generate enough paths in poorly connected topologies in so far as the loopfreeness condition employed (equal cost path, Loop Free Alternate or Loop Free Invariant) is stricter than necessary. In this article, we propose a multipath routing scheme able to compute more loopfree paths (with a low complexity algorithm such as Dijkstra in the worst case and a light communication protocol between directly adjacent nodes) than with existing propositions.