Date of this Version
Recently, many efforts have been undertaken to reduce the energy consumption of core networks. Bundle link is a commonly deployed technique in core networks to combine several high-speed physical sublinks into a virtual connection to achieve bandwidth upgrade flexibility and network reliability. The traffic passing through a bundle link can be carried fully over the first few sublinks (bin packing) or evenly distributed over all sublinks (load balancing). In the current network when a bundle link is on, all of its sublinks are on, thus, selectively shutting down a few sublinks during periods of low traffic could save a large amount of energy while keeping the network topology stable. Previous green network research studies focused on centralized global-optimization techniques which intend to concentrate traffic into a small set of network nodes or links and shut down the other ones under the control of the network management system. Thus, they require frequent changes to the network topology and their solutions are not scalable even with the help of simplified heuristics.
We propose distributed local-optimized algorithms based on thresholds for both bin packing and load-balancing cases to dynamically adjust the number ofactive sublinks. In our algorithm the core routers rely on the link utilization during the previous time slot and use a threshold to trigger the sublinks' up or downoperations. For each bundle link we always retain at least one active sublink to keep the network topology stable. We simulate an Internet2 based syntheticnetwork using bundle links and conduct experiments for both bin-packing and load-balancing cases. The experiment results show that a great deal of (up to86\%) energy consumed on core router ports could be saved with appropriate parameter value settings in both cases. Employing different parameter settingsfor different types of bursty links could greatly reduce congestion with limited loss of energy savings. Compared to previously proposed ILP (Integer linearprogramming) based centralized algorithms, our distributed algorithms can achieve high energy savings and result in fast, autonomous, topology-invariant and scalable solutions.
Adviser: Byrav Ramamurthy