Characterizing nature and location of congestion on the public Internet

Abstract

We address the following question in this study: what is the nature of Internet congestion and where does congestion really occur on a public Internet path? Answering this question helps service providers and content providers better engineer emerging services on the Internet. Our large-scale path measurement and analysis study indicates that congestion on the Internet exhibits a wide variety of packet loss and delay characteristics. Based on our classification using "congestion signatures", we find four dominant "types" of congestion which may be related to macroscopic behavior. A particularly frequent type of congestion we observe is "flash congestion", which creates significant bursty packet loss on a fairly long time-scale. Additionally, our study suggests that flash congestion predominantly occurs at the access provider network within the "last mile" on an Internet path. The Internet "cloud" does not contribute heavily to congestion. Consequently, prevalent approaches to bypass the cloud using "edge-based" content delivery networks and caching may not be effective in reducing congestion.