Load-balancing routing in software defined networks with multiple controllers

Abstract

Software defined networking is a new paradigm that separates the control plane from the data plane. To provide high scalability and reduce the controller load in a large-scale software defined network (SDN), a natural way is to deploy a cluster of distributed controllers so as to cooperatively manage the network. However, since each controller just manages a set of the connected (or associated) switches and usually holds the information of these switches (and connected links), it may result in controller load imbalance and link load imbalance. Thus, both controller load and link load should be optimized to achieve better QoS in SDNs. To this end, this paper tries to answer the following question: how to perform both controller load balancing and link load balancing in an SDN? We formulate the load-balancing routing for both links and controllers (LBR-LC) problem in an SDN, and prove its NP-hardness. A rounding-based algorithm is proposed to solve this problem, and the approximation performance is also analyzed. Moreover, we discuss the efficient mechanism for network status maintenance among distributed controllers. The extensive simulation results show that our proposed algorithm can reduce the maximum controller response time by 70% compared with the previous solution, while only increasing the maximum link load by 3%.

Introduction

Software defined networking (SDN) separates the control plane and data plane on independent devices [1]. The controllers can provide centralized and flexible control by installing forwarding rules in the data plane, and the switches perform different operations on packets according to these rules. With its support of flexible network management, there is an increasing interest in deploying SDNs in different fields, such as data centers [2] and wide-area networks [3].

With the growth of SDN topology, the control plane needs to manage more and more switches, and deal with more and more flows. As specified in the OpenFlow standard [1], current solutions rely on reporting the header packet (or the first packet) of each new-arrival flow to a centralized controller that reactively installs the forwarding rules on switches [4], [5]. If there is only one controller in the control plane, due to its limited processing capacity, it may take a long time or even cause congestion in the control plane. As a result, the single controller may become the bottleneck of an SDN, which will significantly degrade the user experience. To decrease the controller overhead, some researches [6] propose to deploy default paths for all flows using wildcard rules. When a flow arrives at the switch, the switch can find a matched rule for this flow, and forward this flow directly. However, in many practical applications [7], network operators are expected to specify fine-grained (or per-flow) policies that drive how the underlying switches forward, drop, and measure traffic. Since widecard rules only provide coarse-grained flow control, deploying default paths for all flows is not attractive.

Therefore, to avoid such single controller congestion/failure [8], the control plane is typically implemented as a distributed system with a cluster of controllers (e.g., DIFANE [5], NVP [9], and Onix [10]), also called the distributed control plane. There are two main advantages by deployment of multiple controllers. On one hand, when one controller fails, other controllers can still manage the network and determine route paths for flows, which will enhance the network reliability and robustness. On the other hand, since a cluster of controllers can cooperatively process all the new-arrival flows, each controller is responsible for a partial set of new-arrival flows, which helps to reduce the computational complexity and response time of controllers in an SDN.

One key challenge in the distributed control plane is the potential controller load imbalance caused by traffic dynamics. Specifically, the controller may be overloaded if a large number of flows arrive at the switches that are associated with this controller, while in the meantime other controllers may be under-utilized. In the practical networks, traffic dynamics will happen if some applications are generating flows from certain portions of the network, or some switches are serving a large quantity of flows compared with other switches [11]. To this end, it is necessary to handle controller load imbalance.

To conquer this challenge, one way is to permit each switch to change its connected controller from the original one to the target one dynamically, also called switch migration [12], [13], [14], [15]. Though switch migration contributes to controller load balancing, there exist some disadvantages. First, since there is no native mechanism provided in the OpenFlow standard, the switch migration mechanism is not supported by the current OpenFlow standard [12]. Second, some researches try to migrate switches with the help of three operational modes in OpenFlow 1.3, master, slave and equal [13], [14]. This method requires the target controller to obtain enough information about the new switch from another controller, which may take a long time. Since the cost of switch migration is high (long latency, complex migration protocol [14]), it is not suitable for frequent manipulation. Therefore, this paper adopts the static switch-controller assignment, that is, a switch will be always connected with a fixed controller.

In a statically switch-controller configured SDN, aforementioned traffic dynamics will not only pose a severe challenge to controller load balancing, but also unavoidably affect the link load in the data plane. Thus, it is necessary to achieve load balancing on both control plane and data plane. This paper tries to answer the following question: how to jointly perform controller load balancing and link load balancing in a statically switch-controller configured SDN with multiple controllers? What’s more, our key insight is that routing decision will affect the controller load beside the link load. Specifically, we determine the route paths for flows while guaranteeing (1) less traffic on heavy-utilized links, and (2) less computing overhead on heavy-loaded controllers. It is worth noting that these two objectives do not conflict with each other (more details in Section 3.2).

This paper studies the load balancing routing for both links and controllers (LBR-LC) problem in SDNs, and proves its NP-hardness. A rounding-based algorithm is proposed and its approximation performance is analyzed. Moreover, this algorithm is supported by the current SDN standard, OpenFlow, which improves the applicability. We then propose an efficient mechanism for network status maintenance to enhance our proposed routing mechanism. Extensive simulations are conducted to validate the efficiency of our proposed algorithm and the simulation results show that our algorithm can largely reduce the controller response time compared with the existing solutions, while achieving similar performance of link load balancing.

The rest of this paper is organized as follows. Section 2 reviews the related works. Section 3 introduces the problem description and formulation. In Section 4, we propose an efficient algorithm to deal with the LBR-LC problem, and analyze the approximation performance. The simulation results are given in Section 5. We conclude the paper in Section 6.