A Parallel Virtual Queue Structure for Active Queue Management

Abstract

The Adaptive RED proposed by Feng is shown to have small packetdelay and queue length variation for long-life TCP traffic suchas FTP connection with a large file size. However, a greatportion of Internet traffic is short-life web and UDP traffic.Most web traffic has a small file size and its TCP session ismainly operated in the slow start phase with a small congestionwindow size. Since the file size is small, dropping short-lifeTCP (and UDP) packets is not very effective in alleviatingcongestion level at a bottleneck router. From the viewpoint ofTCP, one or several packet losses in its slow start phase lead toextra delay for retransmission and even cause TCP timeout. Thisdelay severely degrades the performance of delivering shortmessages such as web pages and web browsers experience a longwaiting time even with a high speed network. We first show that the Adaptive RED is vulnerable tothese short-life TCP traffic and propose a virtual parallel queuestructure as a new active queue management scheme (AQM). The ideais to separate the long-life and short-life (including UDP)traffic into two different virtual queues. The first queue is torun the drop-tail policy and work for the short-life TCP and UDPpackets. In order to have a small mean delay, the service rate ofthis drop-tail queue is dynamically determined by its virtualqueue length. The remaining long-life traffic is directed to anAdaptive RED virtual queue. Even the available bandwidth isshared with the drop-tail queue, the simulation results show thatthe queue length variation of the RED queue is still located in adesired region. Note that both virtual queues share the samephysical buffer memory. Those packets in the drop-tail queue willnot be dropped unless the shared buffer is overflowed. Thisparallel virtual queue structure not only keeps the benefits ofRED such as high utilization and small delay, but also greatlyreduces the packet loss rate at the router.