Internet congestion control: Cooperative end-system and gateway algorithms.

摘要

In this work, we study the control of Internet congestion from the perspectives of both the network and the end-system. More specifically, we investigate the robustness of the ubiquitous Transmission Control Protocol (TCP) in tbe context of today's Internet traffic characteristics. Secondly, we consider conventional congestion control methods based on gateway packet discard. We make three observations. First, TCP becomes increasingly fragile when the packet loss rate rises, and no previously proposed loss recovery improvements will remedy this problem. Second, to assure TCP's fair and steady delivery of data, the number of network packet buffers should be proportional to the number of active competing flows. Third, as the number of competing users grows, inequity becomes more significant unless proper flow buffer management is implemented at the congested network nodes.; Motivated by these findings, we develop a set of congestion avoidance and control strategies exploiting cooperation between TCP and the network packet discard algorithm. TCP should aggressively recover packet losses while responding to network congestion. Gateways must enforce fairness while managing congestion. Following these guidelines, we develop a loss recovery algorithm (called Net Reno TCP), and a network packet discard algorithm (called Flow Random Early Drop or FRED). Existing TCP suffers from frequent retransmission time-outs due to congestion. Net Reno TCP significantly reduces retransmission time-outs. FRED avoids dropping all in-flight packets of a TCP flow and allows Net Reno TCP to completely eliminate time-outs. Our approach, utilizing both Net Reno and FRED, dramatically improves TCP's robustness.; From implementation's viewpoint, FRED achieves fair queueing with a simple First-In-First-Out policy, and scales well with an increasing number of flows without compromising fairness. We have modified TCP and the packet discard algorithm as found in a BSD UNIX kernel so as to incorporate Net Reno and FRED. The results of laboratory tests are in excellent agreement with our analysis and simulations.