Papers

2016

2015

2014

The OIDA and the Ethernet Alliance organized a workshop to study 100 Gigabit Ethernet over a single wavelength, for the data center market. In my opinion, it is much too early to be worried about this.

2013

This is amazing stuff. You should read it.

2012

This paper proposes a new scheduling algorithm for fast, microsecond-scale data center circuit switches, called traffic matrix scheduling. An advantage of traffic matrix scheduling over the previous hotspot scheduling is that more traffic can be offloaded to the circuit-switched portion of the data center network, making circuit switches much more useful than previously thought.

One weakness of the Helios paper is that the relatively long switching time of 30ms limited the usefulness of the resulting hybrid network. So we developed our own optical circuit switch that is 1,000x faster than Helios. We demonstrated our switch at the SIGCOMM conference.

2011

While working on the Helios project, we discovered that many of the optical communication devices commonly used in data center networks, such as electronic dispersion compensators (EDC) and transimpedence amplifiers (TIA), are not optimized for quickly recovering after a loss of light. However, we did find that such technology exists and future data center networks could be engineered for high performance when using optical circuit switches.

2010

Traditionally, local area networks, including data center networks, have been constructed using switched Ethernet. A typical 10G Ethernet switch uses 12.5W per port, and cost upwards of $500 per port or more. There is a cheaper source of bandwidth: optical circuit switching. This paper describes how we constructed a data center network using an optical circuit switch and what we had to do to achieve good performance.

This is an overview paper of our recent work and a great place to start when learning about the field of data center networking. It covers: commoditization, merchant silicon, scaling Ethernet, and scheduling TCP flows.

This is a collection of paper summaries and transcriptions from the first WREN workshop.

2009

One of the primary reasons for low server utilization in data centers is that the data center network is oversubscribed (underprovisioned). This is necessary due to the non-commodity (super-expensive) nature of current commercial data center switches. This paper describes how to construct a 3,456-port 10G Ethernet switch out of “merchant silicon”, i.e. switch chips designed for low-end commodity Ethernet switches. It is likely that future data center networks will be constructed almost entirely from merchant silicon, just as modern data center servers contain commodity Intel or AMD processors.

A poster version of the Merchant Silicon paper.

Both Layer 2 Ethernet forwarding (learning bridge, minimum spanning tree, broadcast) and Layer 3 IP routing (longest-prefix matching, routing protocols) have problems scaling to large numbers of servers in the data center. This paper describes a new forwarding mechanism called PortLand designed specifically for data centers with much better scaling properties than either traditional Layer 2 or Layer 3 forwarding. Now, at least from a network address viewpoint, it is possible to treat all hosts in a data center as interchangeable components. Additionally, a PortLand network will never contain a forwarding loop.

This was my “Research Exam”, an in-depth survey of a specific area leading to a Masters degree. I have always been fascinated by how simultaneously good and bad TCP is across a wide range of networks and communication patterns. One design feature of TCP is that duplicate acknowledgements are used as an early indicator of packet loss, and hence an early indicator of congestion. Unfortunately, packet reordering will cause the receiver to transmit duplicate acknowledgements, fooling the sender into thinking that the network in congested. This limits opportunities for parallelism in data center networks, which by design have multiple paths between hosts for both increased capacity and fault tolerance. This report looks at historical approaches for fixing TCP so that eventually we will be able to fully utilize multipath data center networks.

2005

Before going to grad school, I worked for the Navy as a robotics research engineer.

2004