There are useful and quite general techniques that implementations can take advantage of for both IGP and MPLS convergence. Prior discussions at NANOG have focused on incremental SPF, but other techniques can be used as well. For example, a two-stage forwarding can allow the IGP route to change and the BGP routes that depend on it to follow with minimal changes to hardware forwarding information.

Tradeoffs exist between using only an IGP vs. using MPLS. The IGP SPF takes order(LlogN), which is further reduced by incremental SPF. MPLS presents scaling limitations with respect to convergence due to the larger number of Constrained Shortest Path First (CSPF) computations that may be required if the set of unique constraints differs. These problems can be avoided through MPLS implementation techniques and network design techniques.

Fast convergence for an IGP is a means to acheive faster restoration when a fault occurs. MPLS has restoration capabilities worth considering. MPLS fast reroute allows fast convergence with a complexity cost in terms of a larger number of LSPs being required. Standby LSP can support subsecond recovery with a bit less complexity. Rerouting LSPs from ingress can be done in seconds for most topologies. This latter case is where MPLS scaling issues come into play. SPF results caching and incremental CSPFs are among the techniques that can aleviate scaling problems, but these too have limits.

There are problems to be solved to make incremental CSPF practical, such as finding similar CSPF results and quickly determining which links would differ for the purpose of incremental CSPF. Adjusting a CSPF for the current path of an LSP when considering rerouting it and the adjusted CSPF needed for disjoint paths present other complexities for which there are solutions.

Topology and protocol usage also affects scaling on the IGP and of MPLS. Implications of area size and LSP tunneling are discussed.

About the Presenter
Curtis Villamizar has been involved in Internet operations, and protocol design and implementation since working for ANS in 1992-1997 in support of the NSF-funded T3-NSFNET project and later ANSNET. In 1997-1999 he was part of the UUNET Network Architecture Group. In 1999 Curtis joined Avici, where he is presently Principal Design Engineer and responsible for Avici's MPLS/TE implementation.