Designing data center networks for high throughput

Data centers with tens of thousands of servers now support popular Internet services, scientific research, as well as industrial applications. The network is the foundation of such facilities, giving the large server pool the ability to work together on these applications. The network needs to provide high throughput between servers to ensure that computations are not slowed down by network bottlenecks, with servers waiting on data from other servers. This work address two broad, related questions about high-throughput data center network design: (a) how do we measure and benchmark various network designs for throughput? and (b) how do we design such networks for near-optimal throughput? The problem of designing high-throughput networks has received a lot of attention, with multiple interesting architectures being proposed every year. However, there is no clarity on how one should benchmark these networks and how they compare to each other. In fact, this work shows that commonly used measurement approaches, in particular, cut-metrics like bisection bandwidth, do not predict throughput accurately. In contrast, we directly evaluate the throughput of networks on both uniform and (heretofore unknown) nearly-worst-case traffic matrices, and include here a comparison of 10 networks using this approach. Further, prior work has not addressed a fundamental question: how far are we from throughput-optimal design? In this work, we propose the first upper bound on network throughput for any topology with identical switches. Although designing optimal topologies is infeasible, we demonstrate that random graphs achieve throughput surprisingly close to this bound -- within a few percent at the scale of a few thousand servers for uniform traffic. Our approach also addresses important practical concerns in the design of data center networks, such as incremental expansion and heterogeneous design – as more and varied equipment is added to a data center over the years in response to evolving needs, how do we best accommodate such equipment? Our networks can achieve the same incremental growth at 40% of the expense such growth would incur with past techniques for Clos networks. Further, our approach to designing heterogeneous topologies (i.e., where all the network switches are not identical) achieves 43% higher throughput than a comparable VL2 topology, a heterogeneous network already deployed in Microsoft’s data centers. We acknowledge that the use of random graphs also poses challenges, particularly with regards to efficient routing and physical cabling. We thus present here high-efficiency routing and cabling schemes for such networks as well.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2016

Author dc:creator

• Singla, Ankit

Contributors dc:contributor

• Godfrey, Philip B
• Gupta, Indranil
• Nahrstedt, Klara
• Shenker, Scott

• topology design
• topology
• network
• random
• throughput

Statement dc:rights

• Copyright 2015 Ankit Singla

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/88997

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/88997