Efficient buffer management for TCP services

TCP suffers from known problems in constrained environments such as satellite and wireless networks. And the problems grow more severe as networks grow increasingly integrated and diverse, where local optimization fail to suffice. One established method for dealing with these problems involves subdividing, or splitting the end-to-end TCP connection into a number of sub-connections. In this paper, we take an indepth look at the buffer requirements for "split connections". After briefly introducing the mechanics of a split connection, we proceed to identify a simple analytic model describing the behavior of the split connection and develop a number of design "rules of thumb" based on our analysis. We compare the results of simulations with the model and measure the efficacy of our design constraints, finding them to improve both the stability and performance of splitting. Finally, we discover that when used properly, oversubscribing, i.e., advertising non-existent space, may greatly reduce the required per-connection buffer-space in some situations.