Flow-length dependent congestion control
No Thumbnail Available
School of Electrical Engineering | Master's thesis
 + 55
AbstractNetwork conditions and traffic requirements of the Internet have been dynamically changing. These changes arise from varying application and user demands from basic web browsing to peer-to-peer applications, server farms, satellite navigation, telecommunication networks, ad-hoc networks and others. Various congestion control mechanisms have been developed to optimize the performance of TCP under such different conditions. This work provides a new TCP congestion control mechanism for supporting the acceleration of shorter TCP flows over longer ones. Two primary motivations exist for such implementation. Firstly, more than 95th percentile of Internet traffic flows is short (< 2MB). Mobile users tend to have even more short flows. Secondly, for wireless devices, the energy consumption is directly proportional to the amount of time their radio interfaces remain open for transmission. Good amount of energy can be saved by finishing flows faster. Our proposed congestion control algorithm is more aggressive than the traditional algorithms at the beginning of the flow and less aggressive if the flow continues for long period of time. The algorithm switches from the aggressive mode to the laid back mode after a threshold is crossed. This threshold has been inferred from the current Internet statistics. Algorithm is proven to accelerate small flows without significantly affecting the longer ones. All analysis is conducted with NS-2. We also have a mathematical model for a similarly behaving algorithm which adjusts the aggressiveness dynamically depending on the flow-length. Aggressiveness of the congestion control algorithm is set to be a function of flow-length in bytes. We are able to finish shorter flows almost twice as fast than normal TCP without causing significant disadvantage to longer flows. Thus in this thesis, we describe the motivation for such algorithm, design two solutions and perform various experiments to see the implications of our concepts.
Thesis advisorSiikavirta, Sebastian
congestion control, congestion pricing, aggressive TCP, TCP fairness, flow-length dependent TCP, internet