Prediction-based dual-weight switch migration scheme for SDN load balancing

Abstract

Software-defined networking (SDN) is considered one of the most promising development modes of the future Internet because of advantages such as programmability and centralized administration. A single centralized controller may cause reliability and scalability issues. Although multiple controllers can solve a single centralized controller${}^{\prime }$s scalability and reliability problems, a flexible mechanism to balance the load is needed. Traffic loads between controllers can easily lead to unbalanced load distribution between them. For multiple distributed controllers, a prediction-based SDN load balancing dual-weight switch migration scheme is proposed. The scheme considers the past traffic load as historical data to predict the future traffic load. Through predictive technology, we know the time when the controller is overloaded, so that the switch migration operation can be carried out in advance. We also propose a triggered load information algorithm to solve the additional processing and communication overhead of the control plane required for periodic active load information between distributed controllers. Considering the information from the past, the proposed scheme suggests that the management of specific switches be migrated between the controllers. We consider the historical load and future load of the switches and propose a switch migration algorithm with dual-weight, it reduces the frequency of switch migration. Experiments have proved that this scheme can quickly balance the load between controllers and reduce the number of switch migrations.

Introduction

Traditional IP networks have a wide range of applications, but they are exceptionally complex and difficult to manage. It is difficult to configure the network according to predefined policies and to reconfigure the network to respond to failures, loads, and changes. Software-defined networking (SDN) architecture is considered one of the most promising architectures of the future Internet. It has a significant impact on the traditional network because of its three characteristics forwarding control separation, centralized management and an open interface [1]. It is a reconstruction of the traditional network architecture and plays an important role in future network transformation [2], [3], [4]. SDN changes the current network${}^{\prime }$s control logic (control plane) to separate the underlying routers and switches (data plane) that forward traffic, thus breaking the vertical integration. Furthermore, the network switches are regarded as simple forwarding devices because of the separation of the control plane and the data plane. The control logic is implemented in a logically centralized controller (or network operating system), thereby simplifying operations such as policy execution and network configuration [5], [6], [7].

Due to the rapid development of the network and the increasing scale of the network, a single centralized controller cannot meet the current network status, and leads to potential scalability and reliability problems. Therefore, some researchers have used multi-controller deployment to avoid this bottleneck [8], [9], [10]. However, a key limitation of these solutions is that when the mapping between the switches and the controllers is statically configured, it is difficult for the controller to solve the load problem between the controllers. This limitation can cause network performance to degrade; therefore, it is extremely important to solve this problem. In OpenFlow V1.3 [11], an SDN southbound interface protocol was proposed for the first time to solve controller load-balancing and failover functionally. The SDN controller has three roles: master, equal and slave. Their roles can be changed at any time [12], but only the controller in the master state can manage and configure the switch. It enables dynamic switch migration to be a simple and effective way to balance the load between multiple controllers.

In the current method [13], the overload controller is always defined as the maximum load controller that exceeds the load threshold, and this method cannot effectively guarantee the accuracy of the controller${}^{\prime }$s load approaching the bottleneck of processing capacity. This always results in frequent switch migration, which can cause interruptions to ongoing traffic. To solve the time and frequent switch migration operations required to collect load information regarding other controllers in distributed decision-making, and to make full use of the historical data on traffic. A new scheme is proposed: SDN multi-controller load-balancing strategy based on traffic prediction, by applying the long short-term memory (LSTM) algorithm and using historical data to predict the future load of the controllers. When the controller is predicted to exceed the load in the next stage, the load balancing operation is started in advance to solve the problem of the sharp increase in the controller${}^{\prime }$s response time caused by the large load. Moreover, to solve frequent switch migration operations, we designed a dual-weight switch migration algorithm.

In this study, our focus is to design a dual-weight switch migration algorithm to solve frequent switch migration operations. To the best of our knowledge, this is the new technology to predict future load data through the historical load data of the controller, and know the time when the controller${}^{\prime }$s high load appears in advance. In summary, the main contributions of this work are as follows.

• We propose the use of the deep learning method to predict future load data based on the historical load data of the controller and to know the time when the controller${}^{\prime }$s high load appears in advance.

• We use a triggered controllers synchronization strategy to reduce the control plane${}^{\prime }$s additional processing and communication overhead required for periodic active load information.

• A dual-weight switch migration scheme is proposed to consider the future switch load situation and effectively avoid frequent switch migration.

The rest of this paper is structured as follows. Section 2 discusses related work, and Section 3 briefly describes the architecture. Section 4 describes the design and implementation of this scheme in detail. Section 5 confirms the effectiveness of our scheme from many aspects. Finally, we summarize our findings in Section 6.