Sally Floyd: Questions

• WHERE IS THE CONGESTION IN THE INTERNET?
  What do we know about congested links in the Internet? E.g., where is the congestion in the Internet? On access links to the home or web server? Peering points? Transoceanic links? Wireless and satellite links? Is the congestion on links with high levels of statistical multiplexing, or on links with low levels of statistical multiplexing? How have the congestion points moved around as the Internet has changed over the last ten years? Where do we expect the congestion to be in the future, if anywhere?

• Venkata Padmanabhan and others have a series of papers on Server-Centric View of Internet Performance: Analysis and Implications that look at the question of identifying congested links, or path segments, from packet traces taken at the web server.
  Results include the following:
  "In 45% of the cases, the identified lossy link crosses inter-AS boundaries and has a high latency." A chart shows how many of the identified non-leaf lossy links cross inter-AS boundaries, and how many have high latency. "Only 20% of the lossy links neither cross ISP boundaries nor have a high latency."

• Anagnostakis, Greenwald, and Ryger, On the Sensitivity of Network Simulation to Topology, MASCOTS 2002.
  "We report that a mesurable fraction of packets pass through multiple congestion points." A congestion point is defined not in terms of packet drop/mark rates, but in terms of the distribution of queueing delay. The rest of the paper shows that if you use total aggregate goodput as the metric, then odd things happen in simulation scenarios with multiple congestion points.

• The continent-to-continent packet loss rates on the SLAC web page are for paths that tend to use reasonably well-provisioned transoceanic links from universities or research institutes (e.g., the GEANT network, StarLight, Abilene-AARnet, the KEK network, etc.). The paths from Egypt, India, Hong Kong, South Africa, the Ukraine, and Vietnam seem to use commercial routes, so the country-to-country packet loss rates from these countries are more representative of commercial paths.

• Aditya Akella, Srini Seshan and Anees Shaikh, Evaluating Wide-Area Internet Bottlenecks.
  "We apply our measurement methodology to determine empirically the locations, estimated available bandwidth, and delay of non-access bottleneck links."

• Mbarika, Jensen, and Meso, Cyberspace Across Sub-Saharan Africa, December 2002, CACM.
  This paper doesn't have information about the congestion levels of international links from Africa, but it does have information about the increasing bandwidth of those links.
  "Until recently few of the countries outside of South Africa had international Internet links larger than 64Kbps, but today 23 countries have links carrying 5Mbps or more, and 13 countries have outgoing links of 10Mbps or more."

• Tom Leighton, Improving Performance on the Internet, CACM, February 2009.
  "When it comes to achieving performance, reliability, and scalability for commercial-grade Web applications, where is the biggest bottleneck? In many cases today, we see that the limiting bottleneck is the middle mile, or the time data spends traveling back and forth across the Internet, between origin server and end user."

• Shawn Ostermann's tcptrace analyzes TCP dump files, and outputs a range of information about each connection seen, including throughput, retransmissions, and round trip times.