Four Network Application Insights From Our AT&T/NJIT Hackathon
Feb 3, 2013
We wrapped up our second hackathon co-sponsored by AT&T and Juniper Networks, co-hosted by NJIT at our OpenLab facility in Bridgewater, NJ. We challenged six teams of students to think about how to improve the quality of experience in a streaming video solution, using Junos Space to gain access to low-level router statistics and configuration parameters and drawing on their own creativity to think about delivering on "better."
Our students didn't disappoint, and there were four insights drawn from their quick pitches that give a glimpse of what's possible in an SDN architecture where you decouple services and their configuration and control.
Algorithm convergence matters. All of the student demonstrations highlighted the phase shift between a change in configuration and the resulting overall throughput and experience seen by users. When those configuration changes are made en masse, there's a risk of oscillation or over-correction - operating on incomplete or partial information, changes are made that may be optimal in a small use case but less so seen in a larger context. One of the benefits of greater centralization is that decisions can be made with more data, from more control points, resulting in less latency to execute and settle into an improved user experience.
There is never just one player. Junos Space gave the students access to control the per-port bandwidth available, and statistics on ingress and egress packet counts. They also had to think about how TCP congestion control algorithms would interact with changes they made in router configuration, and to look at options for controlling the quality and encoding of a video stream to reduce demand on the network. Simply throttling network bandwidth didn't deliver ideal or deterministic results - control and configuration algorithms had to be able to interact with all of the control points from layer 2 through layer 7.
There's no "perfect' answer. When dealing with bursty traffic and a highly stochastic set of network events, you can't grab a perfect answer out of the air - the values that define "perfect" change faster than the inputs that determine improved quality. Moving past purely trailing indicators made a big difference, and those teams that utilized predictive algorithms saw their variability and configuration-to-experience latency improve. That's another reason to think about SDN control and management in a more global, less network element-local way: you can put leading indicators into play.
Draw on other domains.We saw a set of moving average calculations more typically seen in a technical stock market trading application to help bound their predictive algorithm. Students drew on evolving TCP congestion control research, as well as historical facets of ATM network behavior to drive their management solutions. One of the side effects of SDN will be to expand the set of control and management applications to pull from other physical sciences, statistics, behavioral analytics and obviously the emergent domain of "big data."
Working with students always makes you look at problems with a new blend of healthy skepticism, eager curiousity and sometimes a few too many light nights - and our Hackathon teams delivered across the board.