A Study on Software Defined Networking for Data Center Networks

Abstract

Software-Defined Data Center (SDDC) is a new paradigm of managing and operating IT infrastructure, where the resources in data center such as compute, storage, and networking are softwarized and delivered as a service to users on demand via application programming interface. Moreover, these resources are managed and controlled by software automatically--this is unprecedented in traditional IT infrastructure in which the infrastructure is typically defined by and tightly coupled with hardware and software. To realize this ideal environment, SDDC encompasses several virtualization technologies in compute, storage, and networking. In this thesis, we are more focusing on networking because the first two technologies are technologically advanced last few years, but networking evolution is still slow due to the vendor lock-in in which hardware and software of network element, which are typically proprietary, are tightly coupled. Moreover, in the network operators' perspective, configuring box-by-box manner with low-level commands results in increasing management complexity and being error-prone.

To meet the requirements of today's users, enterprises, and carriers, Software-Defined Networking (SDN) is emerged. The main idea of SDN is to decouple control planes from data planes of network elements such as switches or routers and to replace the distributed, per-switch control plane with a (logically) centralized one on which SDN applications can control an operational network with a global network-wide view by enforcing packet forward- ing rules to the distributed data planes. This paradigm shift benefits network operators by (i) reducing the complexity of operations through automation while keeping more responsive, and (ii) optimizing the resources of operational networks with the global network-wide view to meet the dynamic nature of on-demand services in a cloud era. While SDN promises the enormous benefits as mentioned just before, it introduces new challenges: (i) increased control loop--gathers traffic and other measurements from the network and uses the gathered information to compute and install forwarding behaviours in the switches--due to the decoupling, and (ii) limitation on a distributed architecture--a (logically) centralized control plane is horizontally distributed to multiple physical servers--for large-scale production networks due to consistency overhead.

To address the first challenge, we propose, implement and evaluate OpenSample: a low-latency, sampling-based network measurement platform targeted at building faster control loops for software-defined networks. OpenSample leverages sFlow packet sampling to provide near--real-time measurements of both network load and individual flows. While OpenSample is useful in any context, it is particularly useful in an SDN environment where a network controller can quickly take action based on the data it provides. Using sampling for network monitoring allows OpenSample to have a 100 millisecond control loop rather than the 1--5 second control loop of prior polling-based approaches. We implement OpenSample in the Floodlight OpenFlow controller and evaluate it both in simulation and on a testbed comprised of commodity switches. When used to inform traffic engineering, OpenSample provides up to a 150\% throughput improvement over both static equal-cost multi-path routing and a polling-based solution with a one second control loop.

To address the second challenge, we propose FRACTAL, a framework for recursive abstraction of SDN control-plane, to address this problem. In FRACTAL, a large network is divided into multiple small networks, each of which is abstracted as a single virtual switch. This ``divide-and-abstract'' process is recursively iterated until a divided network can be handled by a single controller. A virtual switch is controlled by the higher level controller over OpenFlow, so that FRACTAL can coexist with other SDN mechanisms. We first carry out simulation experiments to demonstrate the issues of naive net- work partitioning. We then implement and evaluate FRACTAL with microbenchmark. Testbed-based experiments reveal that FRACTAL (i) adds small delays for non-local messages that cross divided networks, but (ii) achieves superlinearly increasing (control plane) throughput as the number of abstraction levels in the controller hierarchy grows.