Groupe d’études et de recherche en analyse des décisions


Congestion Avoidance with Future-Path Information


In this paper we analyze the trade-off between admission costs and receiver rewards of TCP Tahoe flows competing for buffer space. Since the buffer space is a scarce resource during heavy traffic and congestion epochs, it is important to understand in which circumstances packet dropping may be optimal. We develop a restless bandit model for assessing an economic value of packets at routers they encounter in transmit. We then argue that the economic value of the whole network increases if the packets with lower economic value are the preferred candidates for dropping or marking in congestion avoidance mechanisms. Such changes are arguably expected to lead both to a lower delay and higher network throughput.

, 16 pages