Delimitation of Macroflows in Congestion Control Management Using Data Mining Techniques

Some of the newest approaches in Internet congestion control management suggest collaboration between sets of streams that should share network resources and learn from each other about the state of the network. A set of such collaborating streams is called a macroflow. In classical congestion control approach, a stream learns information about the network state by itself. It makes use of the acquired knowledge to adapt its transmission rates to the current network conditions. Stream collaboration, in exchange, permits dissemination of network state knowledge: the streams in a macroflow maintain common information about the network state. Every stream in that macroflow uses this shared knowledge and contributes to its maintenance. This dissemination of network knowledge conducts to a better, faster and more flexible adaptation of flow behavior in presence of network congestion. The remaining problem is how to identify the streams forming such a logical entity (a macroflow). Currently a macroflow is organized on host pair basis. We propose in this paper a new method for grouping streams into macroflows if they behave similarly. A flow behavior is described by a set of state variables, such as the congestion window size, the round trip time or retransmission time out. This extended macroflow granularity can be used in an improved Congestion Manager.