Micro Load Balancing in Data Centers with DRILL




Dr. P. Brighten Godfrey

Session Description

A key question for open networking is what functionality should be in network hardware, vs. at the edge or in a controller.  In the sphere of data center networks, the trend towards simpler network fabric strips most network functionality, including load balancing capabilities, out of the network core and pushes them to the edge. We investigate a different direction of incorporating minimal load balancing intelligence into the network fabric and show that this slightly smarter fabric significantly enhances performance. We provide a very simple in-network load balancing scheduling algorithm called DRILL which is purely local to each switch. DRILL leverages local load sensing and randomization concepts to distribute load among multiple paths. Simulations show that this new “micro load balancing” approach can achieve dramatically lower flow completion time and queuing variability than past designs.  This is joint work with Soudeh Ghorbani, P. Brighten Godfrey, Yashar Ganjali, and Amin Firoozshahian.