Speaker: Prof. Brad Karp
Professor of Computer Systems and Networks
University College London, Department of Computer Science
Host: Wenjun Hu
Title: How Topology and Routing Design Determine Internet Backbone Latency
In the ARPAnet and Internet’s early days, we defined “success” in routing as finding some communication path between two hosts despite link failures. As the Internet’s user base and total traffic volume grew rapidly, and applications grew in their capacity demands, the bar for successful routing rose to encompass choosing paths (in many cases, multiple ones) with sufficient capacity to carry users’ traffic. Today, many of the most widely used networked applications offer the best user experience when the Internet not only provides adequate capacity, but also delivers data with *low delay*.
How achievable is low-latency delivery of traffic on Internet Service Providers’ (ISPs’) backbones? In this talk, I’ll explore two distinct yet interestingly intertwined aspects of this question: (1) how well suited are ISPs’ backbone topologies to delivering traffic with low latency?; and (2) how well do state-of-the-art routing systems do at *using* the low-latency paths available in today’s ISP backbones? I’ll present experimental results that show that today’s routing systems (such as Google’s B4) have trouble using low-latency paths effectively in *precisely* those ISPs’ backbones with the greatest potential to deliver traffic with low latency. I will then outline a new approach to low-latency routing that can take full advantage of an ISP topology’s low-latency paths. Along the way, I will explore the fundamental tension in the ISP setting (where the ISP cannot perfectly predict users’ traffic demands) between routing traffic with low delay and risking driving the network into a state of congestion. I will propose link *headroom*, a mechanism that mitigates the risk of congestion in practice. I’ll conclude by briefly considering future prospects for low-latency-capable ISP topologies.
(Joint work with my UCL colleagues Nikola Gvozdiev, Stefano Vissicchio, and Mark Handley.)