Cognitive optical network architecture in dynamic environments

Abstract

Emerging network traffic requires a more agile network management and control system to deal with the dynamic network environments than today's networks use. The bursty and large data transactions introduced by new technological applications can cause both high costs and extreme congestion in networks. The prohibitive cost of massive over-provisioning will manifest as huge congestions during peak demand periods. The network management and control system must be able to sense the traffic changes and reconfigure in a timely manner (in tens of milliseconds instead of minutes or hours) to use network resources efficiently. We propose the use of cognitive techniques for fast and adaptive network management and control of future optical networks. The goal of this work is to provide timely network reconfigurations in response to dynamic traffic environments and prevent congestion from building up.

We make a simplified model of the expected traffic arrival rate changes as a multistate Markov process based on the characteristics of the dynamic, bursty, and high granularity traffic. The traffic is categorized into different network traffic environments by the length of the network coherence time, which is the time that the traffic is unvarying. The tunneled network architecture is adopted due to its supremacy in reducing the control complexity when the traffic volume is at least one wavelength. In the long coherence time regime where traffic changes very slowly, the traffic detection performances of two Bayesian estimators and a stopping-trial (sequential) estimator are examined, based on the transient behaviors of networks. The stopping trial estimator has the fastest response time to the changes of traffic arrival statistics. We propose a wavelength reconfiguration algorithm with continuous assessment where the system reconfigures whenever it deems necessary.

The reconfiguration can involve addition or subtraction of multiple wavelengths. Using the fastest detection and reconfiguration algorithm can reduce queueing delays during traffic surges without over-provisioning and thus can reduce network capital expenditure and prevent wasting resources on erroneous decisions when surges occur. For traffic with moderate coherence time (where traffic changes at a moderate rate) and the short coherence time (where traffic changes quickly), the stopping-trial estimator still responds to the traffic changes with a short detection time. As long as the inter-arrival times of traffic transactions are independent, the algorithm is still optimum. The algorithm provides no prejudice on the exact network traffic distribution, avoiding having to sense and estimate detailed arrival traffic statistics.

To deal with fast-changing traffic, we model the transient convergent behaviors of network traffic drift as a result of traffic transition rate changes and validate the feasibility and utility of the traffic prediction. In a simple example when the network traffic rate changes monotonically in a linear model, the sequential maximum likelihood estimator will capture the traffic trend with a small number of arrivals. The traffic trend prediction can help to provide fast reconfiguration, which is very important for maintaining quality of service during large traffic shifts. We further investigate the design of an efficient rerouting algorithm to maintain users' quality of service when the incremental traffic cannot be accommodated on the primary path. The algorithm includes the fast reconfiguration of wavelengths in the existing lit and spatially routed fibers, and the setting up and lighting of new fibers.

Rerouting is necessary to maintain users' quality of service when the queueing delay on the primary path (determined by shortest path routing) exceeds the requirement. Our algorithm triggers reconfiguration when a queueing delay threshold is crossed on the primary path. The triggering by a threshold on the queueing delay is used due to its simplicity, and it is directly measurable by the exact traffic transaction sizes and the queue size, which reflect both the current network traffic environment and the network configurations. A dynamic rerouting algorithm implemented with a shortest path algorithm is proposed to find the secondary paths for rerouting. We make the conjecture that it is desirable that the alternate paths for rerouting have small numbers of hops and are disjoint with other busy paths when the hops on the path are independent. In addition, the conjecture suggests that a good candidate network topology should have high edge-connectivity.

Wavelength reservation for rerouted traffic does not maximize wavelength utilization. We make the conjecture that traffic with different sizes should be broken up into multi-classes with dedicated partitioned resources and the queueing delay should be normalized by the transmission time for rerouting triggering to realize better network utilization.