Congestion in high-speed wide area networks (HS-WAN) must be controlle
d in order to prevent network performance deterioration, and possible
collapse. Reactive mechanisms applied to slower networks are not entir
ely suitable for HS-WANs due to the relatively long propagation delay
in the latter, On the other hand, preventive mechanisms which have bee
n proposed as possible solutions to this problem assume that traffic c
haracteristics are known at the time of call setup. Window mechanisms
do not depend upon such an assumption, but modifications that have bee
n proposed to adapt it to HS-WANs fall short of the fast response requ
ired in HS-WANs. The Leaky Window (LW) mechanism is proposed in this p
aper as an attempt to solve this problem. This mechanism is a modifica
tion of the sliding window that permits users to transmit traffic in e
xcess of their window sizes based on an estimate of the network load.
The estimate is based on acknowledgments received within a fixed time
interval. Excess traffic is distinguished by 'marking' cells. Marked c
ells are discarded at congested nodes. Congestion control, therefore,
is a local decision executed by the congested node. Through the use of
a simulation model, it is shown that the LW has an average end-to-end
delay and probability of loss that is lower than the sliding window m
echanism. Comparison to the Virtual Leaky Bucket (VLB) mechanism shows
that at lower load (0.7), the VLB has an end-to-end delay and a proba
bility of loss that is lower than that of the LW under the same condit
ions. At higher loads where congestion is a real problem, the performa
nce of the LW is significantly better than that of the VLB.