Summary of results:
We have developed DAMOTE (Decentralized Agent for MPLS Online Traffic Engineering), which provides two main basic functionalities:
- QoS-based routing of DiffServ LSPs (Label Switched Paths) under constraints.
- Local detour (backup) LSP routing for fast restoration.
The first main function of DAMOTE is to compute primary paths at ingress nodes, in a way similar to the classical CSPF (Constraint Shortest Path First). This means that all edge nodes will compute and set up the best path for any given LSP for which they are the ingress. This computation requires that ingress nodes have enough information about all link states in the network. This is usually achieved by using extensions of link-state routing protocols like OSPF-TE or ISIS-TE, which flood the network regularly with updated link-states.
Although similar in principle to CSPF, our scheme generalizes it in several ways. While CSPF is a simple SPF on a pruned topology, obtained by removing links that have not enough resources to accept the new LSP, DAMOTE can perform much clever optimisations based on a network-wide score function. Examples of such functions are: load balancing, hybrid load balancing (where long detours are penalized), pre-emption-aware routing (where LSP reroutings are penalized). DAMOTE is generic in the sense that this score function is a parameter of the algorithm. Like in CSPF, constraints can be taken into account, but here again the constraints can be parameterised quite freely. Typical constraints refer to the available bandwidth on links per class type (CT), or to pre-emption levels.
DAMOTE can also compute local detour LSPs for fast rerouting. In our approach each primary can be protected by a series of detour LSPs, each of them originating at the node immediately upstream of any given link on the primary path. Those detour LSPs thus protect the downstream node (if possible) or the downstream link and merges with the primary LSP anywhere between the protected resource and the egress node (inclusive). Those many LSPs have to be pre-established for fast rerouting in case of failure, and provisioned with bandwidth resource. In terms of bandwidth consumption, this scheme is only viable if detour LSPs are allowed to share bandwidth among themselves or with primary LSPs, which is what we have achieved.
The characteristics of these algorithms are: efficiency, near optimality, genericity, scalability, and pre-emption levels awareness.
Prototyping and testing:
Besides simulations, these algorithms have been implemented in ANSI C to achieve efficiency and portability. The result is a process that has been integrated in Linux and tested on a testbed of 6 Linux PCs configured as MPLS routers. The tested functionality includes primary LSP computation and setup with various score functions, and LSP hard and soft pre-emption and rerouting. The DAMOTE agent is ready for integration in any other platform.
Link with IETF:
Our solutions are aligned with the current work at IETF, and in particular with the Bandwidth Constraint (BC) models of the TEWG. In all cases we rely on the proposed TE extensions of OSPF and RSVP currently being standardized. For some algorithms we need a few additional objects in OSPF-TE and in RSVP-TE PATH and RESV as summarized below.
Application domain:
It should be noted that the overall philosophy has always been to find efficient distributed solutions suited for online traffic engineering. Our solutions are scalable and nearly optimal, despite the fact that the LSPs are established on demand and in sequence. Thanks to its performance, our solutions can thus support the automatic computation and establishment of MPLS-based VPNs at any timescale.
In a hierarchical network where inter-domain LSPs are established, our algorithms can also be used to compute and set up the intra-domain parts of those LSPs.
It is also worth noting that our tools are targeted at strategic TE rather than tactical TE. In tactical TE, LSPs are set up and removed on-demand to circumvent some temporary congested areas, based on some real-time monitoring of the network. In strategic TE, the LSPs are set up to guarantee a certain level of QoS a priori.
However, given the very low computation time of an LSP (usually less than 100ms, even when several class types and pre-emption levels are considered, which remains less than the LSP setup itself), our solutions could also be used for tactical TE. It is even more so because the computation algorithm can take into account the congested links that should thus be avoided, by relying on the existing mechanism that already allows us to compute a new path for a rerouted LSP without reusing the link on which it was pre-empted.
Dissemination:
Our work has lead to 8 papers among which 6 are already published. Our results have been presented at 5 conferences and/or workshops.