Inter-Datacenter Connectivity in Flexgrid-based Optical Networks

The huge energy consumption of datacenters (DC) requires an elastic resource management, e.g. by turning servers off when they are not used or turning them on to satisfy increments in the demand. Thanks to virtualization, jobs (e.g., web applications) can be encapsulated in virtual machines (VM) mixed with other workloads and consolidate them in the most proper server according to their performance goals. Local resource managers in DCs can migrate VMs from one server to another looking for reducing energy consumption while ensuring the committed quality of experience (QoE). Additionally, cloud providers can create DC federations based on a geographically distributed infrastructure so they can manage appropriately green energy resources available in each DC, thus reducing energy expenditure. Scheduling algorithms can perform VM migration not only within a single DC but also transferring a huge amount of raw data from one DC to another to minimize operational costs while ensuring the QoE. Since traffic between DCs is generated by VM migration, the connectivity required between two DCs highly varies along the day, presenting dramatic differences in an hourly time scale. Therefore, using a flexgrid-based optical network to interconnect DCs is an option to be considered since that technology provides fine and multiple granularity. In flexgrid optical networks the available optical spectrum is divided into frequency slices of fixed spectrum width. Optical connections can be allocated into a variable number of these slices, and its capacity can be dynamically managed by allocating or releasing slices provided that the spectrum allocated to an optical connection remain contiguous. Network providers can facilitate the interconnection among federated DCs by allowing them to request connections’ set up on demand with the desired bitrate, while tearing down those connections when they are not needed. With this aim, in the last years, huge standardization work has been done defining control plane architectures and protocols to automate connection provisioning. The Internet Engineering Task Force (IETF) is defining the Application-Based Network Operations (ABNO) architecture, which is based on standard components such as the active stateful Path Computation Element (PCE). This thesis is devoted to characterize, evaluate and analyze the problem providing optimal VM placement so as to minimize operational costs assuming that those costs are dominated by energy and communication costs. To this aim, analytical models to optimize energy consumption in DC federations are provided. Both cloud and core optical network control architectures are explored and new connectivity models for elastic operations are proposed. Mixed integer linear programming models as well as heuristic algorithms are developed and simulations are carried out. More specifically, the main objective has been attained by developing three goals covering different open issues. First we propose the Elastic Operations in Federated Datacenters for Performance and Cost Optimization (ELFADO) problem for scheduling workload and orchestrating federated DCs. A distributed and a centralized approach are studied. Second we propose architectures based on ABNO, using cross-stratum orchestration and carrier SDN, as well as elastic connectivity models supported: the dynamic elastic model and a transfer mode model respectively. Finally, we consider the centralized ELFADO and both the dynamic elastic and transfer mode connectivity models proposed and evaluate their performance