About

Introduction and Motivation

In the crisis of the coronavirus pandemic, the Internet is facing unprecedented surges of traffic induced by the use of cloud-based tele-communting and remote education tools for real-time video meetings and online classes. While sophisticated traffic engineering (e.g., SWAN, B4) enables cloud providers to dynamically reroute traffic to mitigate congestion, interconnections between ISPs and cloud providers (cloud-ISP links) rely on peering links at IXPs, private peering links, or transit providers. These interconnections have limited capacity, and are prone to congestion.
Microsoft recently reported a 775% increase in Teams' online meeting users in a one month period in Italy due to social distancing, and the application of higher bandwidth constraints to cloud applications to mitigate network congestion. This pandemic induced shift in network traffic and amplified the concerns regarding network congestion between these critical infrastructures, as the degraded performance and quality of experience (QoE) of cloud-based applications could negatively impact work and learning performance. However, these is no scientific study to focus on measuring the performance, reliability and resiliency of these cloud-ISP links.
Measuring from cloud providers, such as Amazon EC2, Google Compute Engine (GCE), and Microsoft Azure, can better reveal the performance impact of congestion on cloud-ISP links than existing throughput measurements from the edge, due to the location of available vantage points. For example, M-Lab nodes mostly connect to transit ISPs (e.g., Level3 and TATA). Results from NDT only measure paths from users to their transit ISPs (dark grey line). In contrast, web-based speed tests (e.g., Xfinity speed test and Ookla speedtest) and FCC MBA boxes, which aim to measure access link capacity, often use test servers geographically close and in the same ISP as users, to achieve low latency and maximum throughput. Thus, such measurement traffic often stays within the same ISP (light green line). Traffic induced from these two types of measurements do not traverse the cloud-ISP links (red lines).