Packet based load sharing schemes in MPLS networks

By adding a connection oriented layer to legacy IP routing and forwarding, the MPLS architecture provides new mechanism to perform traffic engineering in IP networks. Current MPLS request for comments and drafts have identified MPLS protection switching and load sharing as attractive applications of these mechanisms, overcoming the limitations of legacy IP routing. Existing load balancing IP routing protocols and architectures (ECMP, OSPF-OMP, etc.) have limited load sharing capabilities to avoid loops (e.g. only for equal cost paths for ECMP), usually based on fixed administrative metrics. Given the connection nature of MPLS networks, arbitrary load sharing algorithms can be performed without leading to loops. In this paper we propose and evaluate a family of load sharing schemes that take into account the nature of the aggregated traffic as well as the link capacities and administrative lengths. We propose cost functions that reflect/penalize LSP congestion, using overflow estimates and we derive optimality conditions. For example, for equal length paths (in the sense of legacy administrative metrics) under particular conditions, the optimal load share is based on the sole bottleneck capacity of the LSP: the optimal load share splits traffic according to the pragmatic rule of proportional capacities regardless of the aggregated traffic, as long as the stability constraints are respected. However, the statistical properties of the traffic (thus the cost functions) give important feedback and stochastic bounds when not in the optimal case, either in the event of a failure or when the real bottleneck capacity differs from the nominal one. We illustrate our schemes with simple but significant examples.