This project addressed the growing challenge of supporting data-intensive scientific research on University campuses.  To improve a researcher's workflow, one can certainly upgrade the campus network to a higher speed, thereby reducing the time needed to transfer the researcher's science data.  However, even with a higher speed campus network research traffic must still compete with general campus traffic for limited bandwidth.  Moreover, conventional campus networks apply the same policies to all forms of network traffic, forcing research traffic to pay the same overheads and pass through the same choke points (e.g., middlebox firewalls, traffic shapers, rate limiters, deep packet inspection boxes, etc.) as non-research traffic.  As a result, research traffic rarely achieves the potential improvements that an upgraded campus network infrastructure should bring.

The typical way to address this problem is to build a special-purpose high speed Science DMZ network outside the University firewalls and then move the University's High Performance Computing system (e.g., the University supercomputer) from the University's production network (a.k.a., campus network) to the new Science DMZ network.  In contrast to the Science DMZ approach, the goal of this project was to transform the University's production network (campus network) into a Software Defined Network (SDN) that could be programmed to carry research traffic at high speed.  In particular, we wanted to use new and emerging SDN capabilities to route research traffic around firewalls and other performance-limiting middleboxes.  Because we are transforming the production network, the new SDN network can provide high-speed not only to the University's HPC systems, but it can also deliver data to/from researchers' desktops and server machines (e.g., file servers and compute servers) all across the campus network at high speed.

Consequently, this project deployed SDN-enabled network equipment across a significant portion of the University of Kentucky campus with the goal of bringing high-speed data transfer capabilities to a wide range of researchers.  In particular, we replaced the existing (production) campus network infrastructure with SDN-enable network infrastructure in several buildings on campus.  We then used network controller software that is able to differentiate between research and non-research traffic, allowing research traffic from any connected system to be routed over high-bandwidth links that bypass middlebox choke points. General non-research traffic on the other hand continues to traverse existing network infrastructure, including policy-enforcing middleboxes.  As a result, a researcher's machine can be transmitting research traffic at high speed (e.g., gigabits per second) while at the same time transmitting non-research traffic at normal speed (e.g., tens to hundreds of megabits per second), with each type of traffic taking an appropriate route across the campus SDN-enabled network infrastructure.  In addition, researchers' machines remain behind, and protected by, the campus security mechanisms by default.  Also, like other software defined networks, network updates and new types of service can be deployed on much shorter timescales.

Our SDN-enabled campus network infrastructure now covers more than a dozen key research buildings on campus, and supports researchers from a wide range of areas including Chemistry, Physics, Biology, Plant and Soil Sciences, Mechanical Engineering, Computer Science, Mining, Statistics, Bio(medical) Informatics, Business and Economics, Communications, and Special Collections researchers.  We compared performance of our new SDN network infrastructure to the previous campus network infrastructure (i.e., non-SDN network) in several of these buildings and found that when transferring data to external Internet sites we would regularly see speeds in the multiple gigabit per second range for SDN, while the previous conventional campus network infrastructure typically operated in the tens or hundreds of megabits per second range.  In other words, we were able to improve researcher file transfer speeds by 1 to 2 orders of magnitude in most cases.  We also worked with researchers to find the best software tools to quickly move data between their desktop/file server and the cloud, and showed that parallel transfer tools could typically be optimized to offer the best performance.

Last Modified: 12/28/2018
Modified by: Brian Nichols