And the OSCARS Goes to ESnet!
Interview With Greg Bell, on Real-World 100G Multi-Layer SDN
This interview by Roy Chua was originally published on SDX Central website.
Greg Bell is director of the Scientific Networking Division at Lawrence Berkeley National Laboratory (Berkeley Lab) in California, and director of the Energy Sciences Network (ESnet), the U.S. Department of Energy’s high-performance networking facility, engineered and optimized for large-scale science. Bell joined ESnet in 2010. Previously, he worked in Berkeley Lab’s IT Division as chief technology architect, reporting to the CIO. Prior to his work for the IT Division, Bell served as a network engineer for Berkeley Lab’s LBLnet Services Group. Before coming to Berkeley Lab, Bell managed a non-profit agency serving political refugees and served as an analyst for Amnesty International. Bell has a Ph.D. in English from U.C. Berkeley and an A.B. in English from Harvard College.
For our readers unfamiliar with ESnet, can you provide them with a brief background?
Bell: Sure. The Energy Sciences Network (ESnet for short) is the Department of Energy’s high-performance science network. ESnet is funded by the DOE’s Office of Science and interconnects the entire national laboratory system in the U.S. We’re different from a commercial provider, because our network is optimized to support very large-scale data transfers. We call these “elephant flows,” in contrast to the smaller “mouse flows” that dominate the commercial Internet traffic mix.
Congratulations on Oscars winning an R&D100 award! Let’s start with the background behind Oscars: What problem were you trying to solve?
Bell: Thanks, we’re very proud of that award. The original idea for Oscars [On-demand Secure Circuits and Reservation System] came from Bill Johnston, former director of ESnet and now a senior advisor to the organization. Almost 10 years ago, Bill noticed that traffic on ESnet was beginning to be dominated by a small number of massive, long-lived flows. Because TCP responds poorly to packet loss when flows are large and round-trip times are long, Bill wondered if we could design software to create dedicated circuits — I call them “fast lanes” — across ESnet and our peer networks as well. The multi-network context is important to us, because large-scale science is highly collaborative.
What is Oscars and how was it built? Who built it?
Bell: Following up on Bill’s initial thinking, a dedicated network engineer named Chin Guok ran with this idea, and the result was Oscars, which is now deployed in over 40 networks around the world. Oscars allows a network operator, end user, or application to create dynamic virtual circuits in real time. These are useful for a number of reasons — for instance, to provide quality-of-service guarantees for particular flows, to achieve security objectives, or to create overlay topologies needed by a large-scale collaboration. I wish we could name and recognize all the organizations that contributed to the success of Oscars. Internet2 has been a critical intellectual partner. Over the years there were important contributions from engineers at the MAX GigaPoP, USC’s Information Sciences Institute, Europe’s Dante/Geant network, Canada’s Canarie, Indiana University, and many others. Oscars has been a great open-source success story.
Bell: In this particular demonstration, the three partners worked together to build an architecture that solves a real-world problem for ESnet. The idea for the demo is to use traffic statistics gathered from OpenFlow and detect one of the massive science-data flows I mentioned, then quickly move that flow down to the optical transport layer, bypassing routing and switching components and optimizing the traffic flow. We used OpenFlow to provision the entire packet-optical path over the Brocade router and Infinera DTN-X switches running Open Transport Switch (OTS) software. Our network load doubles relentlessly every 18 months, and we’re very interested in optimization techniques for controlling costs and ensuring predictable performance.
What’s novel and groundbreaking about this?
Bell: I believe it’s the first demonstration to integrate the detection of large-scale flows with real-time optimization of a multi-layer, multi-vendor network. A demonstration like this goes beyond a simple proof-of-concept and starts to resemble something you’d actually want to build in a production context. It took a lot of work to accomplish, and I’m very grateful to ESnet’s remarkable engineers and to our vendor partners (Infinera and Brocade) for making it happen.
How do you see this applying in the future? Beyond just research networks, what are the implications for service providers and enterprises?
Bell: I think the need to identify targeted flows, then take special measures to forward them differentially, is widely shared. A research network like ESnet may want to provide a specific class of service for flows that use (just as an example) experimental or non-TCP protocols. But this capability should also be useful for service providers concerned with optimizing traffic flows across multiple layers. Of course there are numerous security-related applications.
What do you see as the main benefit of SDN? From your vantage point in working with new technologies, what do you expect to be major trends and changes in networking over the next two to three years?
Bell: For ESnet, the only thing that matters is whether SDN improves the way scientific collaborations get their work done. That’s truly our only metric. I see SDN as a way to bridge the stubborn gap between scientific applications and the network itself. If we can bridge that gap, then applications will be able to make more intelligent decisions based on real-time information from a flexible and programmable network substrate. It doesn’t take a crystal ball to predict that the major trends and changes in networking will revolve around this gap.
What’s your advice to service providers and enterprises around SDN?
Bell: Be empirical. Get your hands dirty. Build a testbed, and convene a team of curious engineers to use it and learn from it. Collaborate with ESnet or another research network on a project that takes us one step forward on the journey towards fully programmable networking.
Greg, it’s been a pleasure speaking with you. Thanks so much for your time and this very informative interview!