Scalable Fast Scheduling for Optical Flow Switching Using Sampled Entropy and Mutual Information Broadcast

Abstract

Optical flow switching is a promising architecture to service large transactions of end users with cost-effective and power-efficient direct access to all-optical networks. For dynamic sessions that are bursty, unscheduled, and only require short transmission times at the full rate of a single wavelength (approximately ones to tens of seconds), the network management and control effort can be substantial, even unimplementable if fast booking and initiation of flow service is needed. In this paper, we describe a fast scheduling algorithm that sets up end-to-end connections for users with urgent large transactions with a scheduling delay of slightly more than one round-trip time. This fast setup of connections is achieved by probing multiple lightpaths between the source and the destination. Probing multiple lightpaths is necessary for moderate to high network loads to achieve low blocking probability. However, the network burden of network-state updates and computational complexity of scheduling can be overwhelming and make the algorithm hard to scale to large networks. With the help of information about network regions periodically updated in the form of sampled entropy and mutual information of the network states, the required efforts can be substantially reduced. To minimize probing efforts and avoid unnecessarily tying up network resources, we use a modified Bellman — Ford Algorithm (Entropy — BF) to select the fewest lightpaths for probing that can satisfy a service-level-blocking probability agreement between the user and the network provider. By collapsing detailsof network states into scalar parameters for average entropy and mutual information, we can greatly reduce both the amount of network-state information collected and/or disseminated and the computation complexity of the lightpath selection process of the probing algorithm. The algorithm is also robust to variations of traffic statistics because it does not depend on detailed assumptions about the statistical model of the traffic, which is often unknown and highly variable in real networks. The throughput performance of this access protocol can be kept high while the network-state protocol burden and computation efforts are reduced by orders of magnitude.