End-to-end transmission control by modeling uncertainty about the network state

This paper argues that the bar for the incorporation of a new subnetwork or link technology in the current Internet is much more than the ability to send minimum-sized IP packets: success requires that TCP perform well over any subnetwork. This requirement imposes a number of additional constraints, some hard to meet because TCP's network model is limited and its overall objective challenging to specify precisely. As a result, network evolution has been hampered and the potential of new subnetwork technologies has not been realized in practice. The poor end-to-end performance of many important subnetworks, such as wide-area cellular networks that zealously hide non-congestive losses and introduce enormous delays as a result, or home broadband networks that suffer from the notorious "bufferbloat" problem, are symptoms of this more general issue. We propose an alternate architecture for end-to-end resource management and transmission control, in which the endpoints work directly to achieve a specified goal. Each endpoint treats the network as an nondeterministic automaton whose parameters and topology are uncertain. The end-point maintains a probability distribution on what it thinks the network's configuration may be. At each moment, the endpoint acts to maximize the expected value of a utility function that is given explicitly. We present preliminary simulation results arguing that the approach is tractable and holds promise.