Thesis Title: Towards Policy-Rich Management of Last-Mile Internet Bottlenecks Radhika Mittal, Chair and Director of Research Klara Nahrstedt Brighten Godfrey Mohammad Alizadeh, Massachusetts Institute of Technology If you wish to attend via zoom, please email Chair or student for Zoom password prior to exam.
Abstract: Billions of users depend on the Internet to exchange data for various kinds of applications. A key resource management decision made on the Internet is how the capacity of any Internet link is shared across flows incident on that link.
Today, in most cases, this decision is offloaded to the congestion control algorithms (CCAs) used by the senders of these flows. Different applications, depending on their needs, may use different CCAs that are unable to compete with each other to fairly share the capacity of the underlying link.
In fact, fairly sharing the capacity may not be a worthy end-goal, as different flows carry different utility to end users. Regardless, it is quite challenging to scalably implement any mechanism to share the capacity of a link across occupying flows either fairly or by any other richer policy.
This is because Internet routers have finite resources (e.g., queues, memory), usually not enough to implement any complex rate-sharing policy reasonably. We note in this thesis that the key bottlenecks faced by most flows on the Internet today are the last-mile artificially rate-limited links.
For example, Internet Service Providers (ISPs) rate-limit users' downlink traffic to their subscribed rate. In this context, we argue that it should be the end-user who decides how their access link bandwidth is shared across their downstream flows.
To enable this, we first present CRAB: a mechanism that allows end-users to directly control how their access link bandwidth is shared across their incoming flows without any support from the senders or the ISP.
CRAB achieves this by throttling flows at the receiver based on the user-specified preferences and estimates of bottleneck link capacity and flow demands. The key challenge CRAB addresses is the accurate estimation of capacity and flow demands while simultaneously throttling the flows.
Next, we present BC-PQP: an efficient and policy-rich rate enforcement mechanism that allows ISPs to directly incorporate rich rate-sharing policies inside their rate-limiting at a minimal cost. Our key insight is to use traffic policers instead of traffic shapers for rate-limiting and extend their capability to support a rich set of rate-sharing policies.
We show how BC-PQP can support a rich set of rate-sharing policies while doing burst-free correct rate enforcement at 7x lower cost than existing mechanisms. Next, with QLDL, we show how traffic policers can be augmented with a light-weight feedback mechanism to enable high-throughput transmissions while incurring zero queuing delay or packet losses.
Our prototype evaluation using WebRTC applications shows how QLDL thus achieves zero queuing and drops, thereby improving interactive applications' QoE on all fronts: 2x higher bitrate, 2-3x lower frame delays, and 20x lower playback freezes, when compared to state-of-the-art baselines.