Clean Slate Research Projects

10.) Rate Control Protocol

Overview

Rate Control Protocol (RCP) is a new congestion control algorithm designed for fast download times (i.e. aka user response times, or flow-completion times). Whereas other modifications to TCP (e.g. STCP, Fast TCP, XCP) are designed to work for specialized applications that use long-lived flows (scientific applications and super computer centers), RCP is designed for the typical flows of typical users in the Internet today. For example, a mid-size flow in the Internet today contains 1000 packets and TCP typically makes them last 10x longer than need-be (XCP is even worse). RCP makes flows finish close to the minimum possible, leading to a perceptible improvement for web users, distributed computing, and distributed file-systems. We believe RCP is the only congestion control algorithm to do this. The main properties of RCP are:

• Typical Internet flows will see 10 times faster download times than TCP and 30 times faster than XCP. Winners are the greater than 90% of sessions that never leave slow-start today.
• Efficiently uses high bandwidth-delay product networks such as the long haul optical links
• Provably stable network independent of link-capacities, round-trip times and number of flows
• Flows are easy to police, to ensure they adhere to congestion control (not generally possible with other schemes)
• Network operators can give preference (or weighted preference) to some flows/aggregates.

 

RCP has two components: (1) End-host congestion control layer that sits between IP and TCP/UDP. During introduction, the end-host could adapt by testing for RCP at each end and along the path, falling back to TCP if need-be. (2) Each router maintains a single fair-share rate per link. Each packet carries the rate of the bottleneck link. For each packet, the router compares the two values. If the router's fair-share rate is smaller, it overwrites the value in the packet. This way, the source learns the fair-share rate of bottleneck link. It is simple, requires a very minor change to switches/routers, requires no per-flow state or per-packet computation.

---