Optimal Qos-aware network reconfiguration in software defined cloud data centers

Abstract

Software-defined networking (SDN) as a new paradigm for networking provides efficient resource reallocation platform in emerging cloud data center networks. The dynamic nature of cloud data center network's traffic, as well as the existence of big flows make it necessary to periodically reprogram the network through the SDN controller. Therefore, it is critical for network researchers to minimize the side-effects of network reconfiguration. In this way, the most challenging issue is the number of rerouted flows that affect the network stability and QoS parameters. As a result, dynamic reconfiguration of the network with minimum overhead (i.e. minimum flow rerouting) is an interesting problem in SDN-based resource reallocation. In this paper, we mathematically formulated the resource reallocation problem as an optimization problem with minimum network reconfiguration overhead subject to QoS requirements of the applications’ flows. In order to reduce the time complexity of solving the optimization problem, a forwarding table compression technique is devised making the proposed scheme an efficient resource reallocation method which can be used as an application on top of the SDN controller. Our Experimental results show that the proposed scheme decreases the network reconfiguration overhead dramatically while meeting the QoS constraints. Since the reconfiguration overhead of the proposed scheme is low, the controller can reallocate the resources more frequently based on the network condition. We also studied the impact of the proposed network reconfiguration scheme on packet loss and delay in the network. The results show that the proposed approach outperform the conventional methods.

Introduction

The emergence of new bandwidth demanding applications such as video, cloud and over-the-top (OTT) services and also smart devices (e.g. Smartphones and Tablets) makes an evolution in computer network's traffic pattern. Cloud computing is developed to support the requirements of these new services and devices. As a consequence, data centers have evolved from a room packed with workstations to large-scale warehouses including hundreds of thousands of servers. In these new large-scale data centers, network infrastructure and protocols play a critical role. On the other hand, the quick acceleration of the Internet evolution has made the weaknesses of its current architecture and protocols so visible that they can be no longer masked by simple over-dimensioning network infrastructures [1]. These issues motivate the network researchers to a revolution from traditional networks to programmable networks. In this way, the state-of-the-art paradigm called software-defined networking (SDN) gives a new perspective to the network management in which the control plane is centralized and provides a global knowledge to the network managers. In brief, to address the mentioned issues SDN (which is a programmable network technology that is proper for dynamic and high-bandwidth network applications) is proposed.

Lots of today network applications such as media streaming, online gaming, cloud services and etc. require predictable, steady network resources with strict QoS requirements. In software-defined networking, OpenFlow provides flow level programmability that can be leveraged to program the network according to QoS requirements of the applications and also network traffic condition, dynamically. As a result, QoS-aware network reprogramming or reconfiguration plays an important role in traffic steering in multiservice SDN-based networks. In this type of resource reallocation, links are assigned to network traffic flows based on their QoS requirements and also traffic engineering objectives. In this way, there are several important issues that make flow level resource allocation more challenging: 1) big flows and resource partitioning in the network, 2) existence of burst and dynamic traffic, and 3) various traffic classes with different requirements. One of the most important side-effects of network traffic dynamicity and big flows is resource partitioning. Suppose there are two 10 Gb/s path from node A to node B. Each of them has 500 Mb/s free bandwidth and there is a flow with 800 Mb/s rate from node A to node B. Due to the resource partitioning, the flow cannot be routed properly, however, the network has 1 Gb/s free bandwidth capacity. In order to overcome the effects of resource partitioning, the flow routing must be done with a global view of the network and also considering its impact on all of the other flows. In other words, in order to route a big flow, we may need to reroute a number of other flows due to resource partitioning. In this way, the static network configuration methods are inefficient. Besides, with the existence of big flows and dynamic traffic in the network, it is necessary to reprogram the network frequently which may result in instability and undesirable impact on QoS of the flows. Therefore, these network characteristics make the problem of network reconfiguration with minimum side-effect and overhead (i.e. minimum changes in forwarding tables) more interesting in SDN networks. It should be mentioned that the network reconfiguration overhead (or precisely the flow rerouting overhead) is dependent on the number of rerouted flows. Increasing the number of rerouted flows not only may result in the network instability but also it may increase the packet loss and end-to-end delay.

In brief, the most important differences of our work with the traditional approaches are as follows: 1) lots of traditional approaches focus on the routing of new flows while QNR (the proposed QoS-aware Network Reconfiguration) focuses on the rerouting of existing flows. 2) Since the proposed algorithm considers the effect of flows on each other, it can handle resource partitioning. 3) Despite traditional algorithm, QNR can be used along with any other routing algorithm. 4) QNR focuses on network reconfiguration overhead and reroutes the flows in a way that minimizes the network reconfiguration overhead.

In this paper, the QoS-aware network reconfiguration problem with minimal changes in the forwarding tables is mathematically formulated. In order to overcome the side-effects of resource partitioning in the network, the proposed scheme is considered as a multi-path approach in which two flows from an identical source and destination can be routed through different paths. Additionally, it reallocates the resources in a way that imposes the minimum overhead to the network in reconfiguration phase. Since there are various traffic classes in the network, the proposed approach supports k different traffic classes with various QoS requirements. Briefly, a dynamic and efficient network reconfiguration scheme is proposed in which flow bandwidth requirements are guaranteed and the network reconfiguration overhead is minimized. Besides, in order to reduce the time complexity of network reconfiguration, we have also proposed an approach to achieve a near-optimal solution. In this way, a forwarding table compression technique is exploited to make the scheme more efficient. Moreover, we made a tradeoff between time complexity and the accuracy of the solution by setting a variable compression rate for the flows. Therefore, the exchanged information between a special application with another application in another server can be considered as a flow while all traffic from one data center to another one can be considered a flow, too.

The main contributions of this paper are summarized as follows:

• Mathematically formulation of the QoS-aware network reconfiguration problem with minimal changes in the forwarding tables.

• Addressing the resource partitioning problem in flow routing.

• Using a forwarding table compression technique to reduce the complexity of network reconfiguration.

The rest of this paper is organized as follows: related QoS-aware resource allocation algorithms are described in Section 2. The assumptions and basic information that are required to define our scheme are discussed in Section 3. Section 4 and Section 5 describes the main problem of this work as well as our solution. Section 6 compares the proposed solutions with a traditional approach. Finally, the paper is concluded in Section 7.