Research On Bulk Transfers And Updates In Software-defined Networks

With the rapid development of technologies such as cloud computing,big data and artificial intelligence,all walks of life are accelerating the process of automation and intelligence in today's digital society.This in turn is placing higher requirements on network performance.However,the traditional IP network is increasingly complicated and struggling to manage and keep pace.In traditional IP networks,the control plane and data plane are tightly coupled in distributed devices,which increases network management difficulties and hinders network innovation.To address these issues,Software-Defined Networking(SDN)technology has emerged,which improves the network flexibility by moving control logics off of underlying network devices to a logically centralized controller.SDN also introduces programmability to the network,which simplifies network management and facilitates rapid network innovation.Many technology companies such as large-scale online service providers are adopting SDN technology to completely revolutionize their network design and operations.SDN has become the trend of network evolution.In spite of the above benefits,SDN still faces fundamental challenges.In this thesis,we study two key issues of SDN,consistent and congestion-free flow updates and the deadline-aware bulk data transfers.The main contributions are as follows.1)Consistent and congestion-free updatesIn SDN networks,updating the dataplane routes is asynchronous across multiple switches,which may cause inconsistent forwarding problem.To address this problem,we propose a mechanism to ensure the forwarding consistency of each flow in the process of updates.Our solution is inspired by the Lego game,which leverages existing routes in switches to act as building blocks and lets packets traverse through multiple building blocks that can concatenate the new routes after updates.Such a mechanism can help the updating flows quickly apply their new routes and avoid the rule duplication in the process of an update.Another challenge of an SDN update is the transient congestion due to the chaotic traffic migration of multiple flow.In solve this problem,we propose to schedule flow updates to avoid congestion that may happen during an update.To this end,we formulate the update scheduling problem as a mixed integer linear programming and develop a simple yet efficient algorithm to find a schedule that can migrate all the flows in the minimum rounds and without transient congestion.In addition,SDN updates may have more diverse requirements,which usually requires network operators to spend significant effort to manage different updates.To simplify the update management tasks,we propose a tool for operators to schedule different updates by using high-level abstractions and generic formulations.Experimental simulations show that our proposed solutions outperform prior work in effectiveness and efficiency.2)Deadline-aware bulk transfersOptimizing bulk transfers with mixed deadlines.We disclose the inefficiency of current traffic engineering practices to schedule bulk transfers with a mixture of hard and soft deadlines.To improve the system utility,we employ a unified revenue model that considers the revenue earned from the deadline-met traffic and penalties paid for the deadline-missed portion.We formulate the bulk transfer optimization problem as a linear programming and propose an efficient online algorithm to maximize the service provider revenue.Meanwhile,we conduct a thorough theoretical analysis and show that the proposed algorithm achieves a competitive ratio to the optimal.Extensive simulation results also demonstrate the effectiveness of the proposed solution.Optimizing bulk transfers with point to multipoint communication pattern.The georeplication service creates an increasing amount of bulk transfers with point to multipoint communication pattern.However,there is little work on guaranteeing deadlines for such traffic.To bridge this gap,we propose a solution that can guarantee the deadline of admitted transfer requests and use the network bandwidth efficiently.The main idea is to let some finished receivers act as sources to accelerate transfer completion by joint application of source selection and spatial-temporal routing and bandwidth allocation.We develop two fast algorithms to schedule transfers with long deadline and many receivers,respectively.The simulation results show that the proposed solution outperforms prior solutions in that it provides deadline guarantees for significantly more transfers,achieves higher network throughput,and sharply reduces the transfer completion time.Cross-layer resource optimization for bulk transfers in reconfigurable networks.At last,we initiate the study of how to optimize bulk transfers subject to point to multipoint communication and strict deadlines,leveraging both SDN-enabled multicasting and reconfigurable topologies.To this end,we propose an approach,named DaRTree(Deadline-aware Reconfigurable Tree).While combining multicasting and topology reconfiguration optimizations,DaRTree outperforms state-of-the-art of approaches already for just one of these optimizations.i)Even under a workload which consists only of point to point transfers,DaRTree outperforms prior work that is based on local search heuristics to reconfigure the WAN,by efficiently relaxing and rounding an integer program formulation.ii)Even if the WAN topology is static,DaRTree outperforms prior multicasting approaches as well by performing load-adaptive routing.iii)The extensive simulations on real-world topologies show that DaRTree also significantly outperforms the state of the art in network throughput and the number of deadline-meeting requests.