The rise of heterogeneous congestion control algorithms and complex application loops has led to growing concerns that network bandwidth allocations are unfairly skewed when internet services are competing over shared bottlenecks. Some services (winners) consistently get more bandwidth than their fair share at the expense of others (losers). This website displays data from the study done under the Prudentia Project. The Prudentia Project studies fairness between live, end-to-end services with distinct workloads under various settings that mimic ordinary usage. The study examines services, complete as deployed, with all their complexities.
Below is a list of services that we measure:
Congestion Control Algorithms
In our experiments, we use iperf to let a congestion control algorithm transmit data as fast as possible. They include:
Video Streaming Services
For the video streaming services, we run a program that aims to acquire the maximum amount of data from the service, simulating a video streaming scenario. Most of these video streaming services are rate limited at some cap meaning that even when the link bandwidth is increased the service would not take any more of it. For example, if a service is rate limited at 10Mbps per setting, if it competes with another service on the same network link of 8Mbps, the services’s max-min fair share would be 4 Mbps (half the shared link). However, if the bandwidth is 50Mbps, then the application cannot claim more than 10Mbps because that’s the maximum it can achieve. Therefore its max-min fairshare at 50Mbps is only 10Mbps, while the competing services get the remainder of the bandwidth (25 + (25-10)) = 35Mbps. The video services include:
Bulk download services
For the bulk file download services, we run a program designed to obtain the maximum amount of data from a file stored in the service, replicating the process of downloading a large file. They include: