Aggregate flows: for efficient management of large flows in the internet

Abstract

Considerable advantages for performance will accrue if large flows can be identified and treated differently in the Internet. DiffServ provides a way to treat flows distinctly in the internet, however configuring DiffServ requires a high level of skill and attention to detail. We seek an approach which is inherently autoconfiguring. In this paper flows are not defined uniquely by information in their packet headers, but instead by their route through the local routing domain. This approach imposes a topological structure on the collection of flow by means of which we can judge how close two flows are together and, where appropriate, judge two or more flows to be part of a single larger flow. The aim is to get closer to the collection of packets generated by a single human request for service. For example, if a user chooses to subdivide a request as a large number of TCP connections, the entire aggregate would still be identified and treated as a single flow. To reduce the complexity of the task of flow management, we identify only flows larger than a certain threshold, without having to deal with the much larger number of small flows. Simulations are used to determine how many large flows need to be found in order to achieve significant improvements in the performance of the Internet as a whole.