Home // ICSNC 2017, The Twelfth International Conference on Systems and Networks Communications // View article

Evaluation of SDN Enabled Data Center Networks Based on High Radix Topologies

Authors:
Bogdan Andrus
Victor Mehmeri
Achim Autenrieth
Juan Jose Vegas Olmos
Idelfonso Tafur Monroy

Keywords: SDN; Data Center topologies; Torus; Hypercube; Jelly Fish; Fat Tree

Abstract:
The relevance of interconnects for large future datacenters and supercomputers is expanding as new technologies like Internet of things (IoT), virtual currency mining, High Performance Computing (HPC) and exa-clouds integrate further into data communication systems. The Data Center Network Layer is the workhorse that manages some of the most important business points by connecting the servers between them and delivering high performance to users. Evolution of the networking layer has seen, in addition to improvements of individual link speeds from 10Gb/s to 40Gb/s and even 100Gb/s and beyond, quite important changes in the topological design. Traffic intensive server-centric networks and high performance computing tasks are pushing a shift from the conventional Layer 2 oriented fat tree architectures with multiples tiers towards clos networks and other highly interconnected matrices. Optimal performance and reliability perquisites imposed on the network cannot be fully achieved by solely changing the topological design. A software-centric control of the network enables the use of additional redundant paths not only for increased performance but also reliability concerns. By decoupling the network control from individual devices and centralizing the network intelligence inside a Software Defined Network (SDN) controller, dynamic workloads can easily be accommodated with the deployment of custom modules or applications for traffic management. In this paper, we focus on the innovations for next generation data center networks from a twofold perspective. On the one hand, we evaluate the applicability of new potential interconnection schemes like torus, hypercube, fat tree and jellyfish in regard to identified key metrics such as performance, complexity, cost, scalability and redundancy. Our evaluation comprises of a mathematical interpretation of the graphs with a focus on the abstract metrics (e.g., bisection bandwidth, diameter, port density, granularity etc.) followed by a simulation of the scalable networks in a virtual environment and subject them to various traffic patterns. On the other hand, we introduce an emulated SDN test framework, which decouples the control plane from the interconnection nodes and gives a centralized view of the topology to a controller handling the routing of the internal workflows for the data center. With the use of our SDN enabled testbed we demonstrate and highlight the clear superior performance gain of centralizing the network intelligence inside a software controller, which allows us to apply a custom routing algorithm.

Pages: 30 to 35

Copyright: Copyright (c) IARIA, 2017

Publication date: October 8, 2017

Published in: conference

ISSN: 2163-9027

ISBN: 978-1-61208-591-3

Location: Athens, Greece

Dates: from October 8, 2017 to October 12, 2017