Modern broadband integrated service digital networks (B-ISDN) must handle multiclass traffic with diverse quality of service (QOS) requirements. The main purpose of our research is to design call rerouting mechanisms which provide rapid restoration of network services in case of link failures. We suggest two approaches: Virtual circuit (VC) and virtual path (VP) reroutings. The first approach is more reactive while the latter is more proactive. The applicability conditions for the first approach include the availability of a layered network structure similar to VC/VP architecture which is widely accepted in asynchronous transfer mode (ATM) networks. Another applicability condition is the extent of network failure: VP level restoration is designed for single link failures - the most common in the telecommunication networks. On the other hand, in case of less predictable multiple link failures, VC-level rerouting is appropriate. These two rerouting approaches vary in the amount of time required to carry them out. Though both schemes are designed to work in real time, VP-level rerouting tends to be faster and can be performed in an on-line mode using pre-computed paths. VC- level rerouting requires real-time computation of routes which may result in a noticeable impact on some services. On the other hand, VP-level rerouting requires a substantial amount of off- line computation to design the VP layout and the backup routes.

In this dissertation we propose a new model and associated algorithms to solve a VC-rerouting problem in real time. This model takes advantage of the distributed network data and computational resources by decomposing the problem at an early stage and then performing the computations in a decentralized mode.

In order to solve the fault tolerant VP layout problem, we formulate a bi-criteria optimization model reflecting the tradeoff between throughput and certain QOS requirements. The model involves a piece-wise linear approximation to the capacity allocation rule for variable bit rate connections statistically multiplexed over a VP.

Both models are formulated as integer programs. The solution method developed employ relaxation and aggregation of variables, feasible solution heuristics and valid inequalities. The results of the computational experiments presented indicate that the methods developed are efficient and produce accurate solutions.