A novel traffic independent NNC for dynamic buffer tuning to shorten the RTT of a TCP channel

Abstract

The proposed neural network controller (NNC) is for user/server level dynamic buffer tuning. It eliminates overflow by making the buffer length always cover the queue length adaptively on the fly. The NNC and the AQM (active queue management) mechanism(s) at the system/router level together form a unified buffer overflow prevention solution for the TCP channels. The NNC provides the following advantages: a) it prevents the AQM resources dished out already at the system level from being wasted, b) it shortens the service roundtrip time (RTT) by eliminating the chance of retransmission due to user-level overflow, and c) it alleviates network congestion. The NNC operates by the {0,/spl Delta/}/sup 2/ objective function and the principle is to maintain the given /spl Delta/ safety margin about the reference point symbolically represented by the "0" in {0,/spl Delta/}/sup 2/. The NNC stability and precision are traffic pattern independent for its statistical nature. This independence makes the NNC immensely suitable for Internet applications because Internet traffic can be LRD (long-range dependence), SRD (short-range dependence) or multifractal, and switch suddenly.