Efficient Traffic Management Of Evolutional Network Architectures

The presence of multiple paths interconnecting the network end points allows operators to use multiple paths simultaneously to avoid congestion when run at high utilization.Thus,network operators typically deploy load distribution strategies in using multiple paths to spread incoming traffic on the multiple paths in the network in order to exploit the high bandwidth.Traditionally,operators used one of the most well-known routing techniques called Equal-Cost Multi-Path(ECMP)routing,which routes each incoming flow to one of multiple paths of equal cost at random.However,traditional network switches have tightly coupled control plane and data plane.As a result,network operators had to manually login to each individual switch to configure them.It is challenging to correctly and effectively implement management tasks on traditional network.Software-Defined Networking(SDN)decouples the control plane from the data plane.The control plane is a logically centralized controller.It gathers information form the data plane and provides a global view to the operator.Thus operator can make traffic assignment decisions based on the global view and configure the data plane via the controller.While SDN is an at-tractive paradigm for operator to achieve flexible routing algorithms like traffic engineering etc,the controller-based centralized routing algorithms are too slow to react to changing network conditions,which often take on the order of minutes.Programmable data planes allow the operator to specify the way of packet processing in the hardware data plane.The most important aspect of programmable data plane is that algorithms can access stateful memory registers that can be used to monitor changing network conditions.Thus multipath routing schemes can react to changing network states on the data planes.How-ever,exploitation of such data plane architectures that allow for stateful packet processing entirely in the data plane is fraught with many challenges.Specifically,operator should firstly decompose network function into multiple functionalities,then design the way in which those functionali-ties interact with each other,and finally implement those functionalities on distributed switches.These often involve error-prone manual function decomposition and network-device configura-tion.Network operators build various network architectures to simplify network management,such as traditional network,SDN,and data-plane programmable SDN,and run various applica-tions on the control plane to perform different management tasks,like load balancing,traffic en-gineering,and flow scheduling.Today,a lot of Internet services,such as social networks,search engines,and e-commerce are hosted in data centers.End users use their personal computers,mobile phones and tablets to access these Internet services via networks in campuses,compa-nies,ISPs(Internet Service Provider),and data centers.Managing these networks to provide fast Internet services is a central problem for computer networking research.Network operators devote tremendous time and effort to three key management tasks,i.e.,traffic engineering,load balancing,and flow scheduling.It is challenging to efficiently perform these management tasks on top of various network architectures.In this paper,we present a new network traffic management framework that can efficiently handle network traffic for multiple applications and across the network architectures.We identify and study the following three key components for the framework.Load Balancing in Traditional Networks:Load balancing can effectively improve net-work performance and scalability,but it may cause packet disorder,so worsening the per-formance.Additionally,without MPLS(Multi-Protocol Label Switching)to establish the desired end-to-end paths,hop-by-hop routing load balancing is more difficult to achieve than the source routing;however,it can significantly improve network performance.In this work,we propose a load balancing scheme with hop-by-hop routing,by using the burstiness features of flows to make sure that the packets of the same flow arrive at the re-ceiving end in order.Simulation results show that our algorithm can adapt to the dynamic changes of the end-to-end delay and the routing vector,and also can achieve fine-gained load balancing.Traffic Engineering in Software-Defined Networks:This work is driven by a simple question of whether traffic engineering in SDN can react quickly to bursty and unpre-dictable changes in traffic demand.The key challenge is to strike a careful balance be-tween the overhead(frequently involving the SDN controller)and performance(the degree of congestion measured as the maximum load and the balance between the minimum and the maximum loads).Exploiting OpenFlow features,quick shift of routing paths for unpre-dictable traffic bursty is the focal point of this work.It is achieved by using a dual routing scheme and letting the data plane to select the appropriate path in reacting to uncertainty in traffic load.The proposed work is called DUCE(Demand Uncertainty Configuration sElection).Further,we describe a traffic distribution model,an optimization solution that calculates congestion-free traffic distribution plan which guarantees that each switch can select one of the paths in a distributed way,and moreover,OpenFlow details about de-taching the functionality of responding to the demand uncertainty from the control plane and delegating it to the data plane.Simulations are performed validating the efficiency of DUCE under various network scenarios.Traffic Management in Programmable Data-plane:Network programming languages have emerged to simplify the programming of stateful packet processing in the data plane of software defined networks,which requires detailed knowledge about the packet processing patterns.However,network functions usually exhibit high level behaviors in terms of traf-fic semantics,path condition,and path computation.To bridge the gap,we propose EP2(Easy Path Programming)that offers a higher abstraction level to simplify the program-ming of network functions.EP2 captures the common properties of network functions and uses predicates and primitives as basic language components.Specifically,predicates are used to describe when to handle a flow by providing programmers a global view of the flow semantics and path conditions;and primitives are used to describe how to choose a path for a specific flow through enabling the centralized control over path computation.Further,EP2 decouples the implementation challenges from programming model and del-egates them to the EP2 runtime system.These implementation challenges include how to decompose high-level function into distributed functionalities,where to place the corre-sponding functionalities,and how to connect these functionalities to achieve the desired behaviors.Throughout the paper,cases are given to illustrate the benefits of EP2 when the language details are introduced.The expressiveness of EP2,the potential overhead of the runtime system and the efficiency of the network functions generated by EP2 are evaluated.The results show that EP2 achieves comparable performance while reducing programming efforts.