Designing the Network Layer for End-Host Traffic Engineering

Traffic engineering (TE) is used to control the distribution of network load across the links and routers of a network. TE techniques usually focus on multiplexing traffic across competing paths connecting a source and destination. There are two significant barriers to traffic engineering on the Internet. First, it is difficult to employ interdomain traffic engineering (across multiple autonomous systems) because BGP forces a router to select a single best route per prefix. Second, intradomain TE is complicated because the paths of individual flows can not be changed arbitrarily without negatively affecting high-layer protocols, particularly TCP. These disparate difficulties can both be addressed by exposing a portion of the TE decision process to end-hosts. We propose a tag mechanism that allows end-hosts to influence, but not control, the paths chosen by the network for its packets. End-hosts will provide explicit opportunities for TE by tagging packets with opaque identifiers. Only packets with the same tag must be delivered along the same path. The tag mechanism will also allow multiple ASs to coordinate to provide interdomain traffic engineering. Our analysis of Internet provider/peering topologies shows that there are multiple paths available between the majority of ASes that do not violate AS transit policies, and that tags would expose these paths to end-host traffic engineering.