3Com Switch 8800 Advanced Software V5 Configuration Guide
65
CONGESTION AVOIDANCE
When configuring traffic classification and traffic shaping, go to these sections for
information you are interested in:
■ “Congestion Avoidance Overview” on page 859
■ “Configuring WRED” on page 861
■ “Displaying and Maintaining WRED” on page 862
■ “WRED Configuration Examples” on page 862
Congestion Avoidance
Overview
Excessive congestion can endanger network resources greatly, so some congestion
avoidance measures must be taken. Congestion avoidance refers to a traffic
control mechanism that can monitor the occupancy status of network resources
(such as the queues or buffer). As congestion becomes worse, the system actively
drops packets and tries to avoid the network overload through adjusting the
network traffics.
Comparing with the end-to-end traffic control, this traffic control herein is of
broader significance, which affects more loads of application streams through a
device. Of course, while dropping packets, the device may cooperate with traffic
control actions (such as TCP traffic control) on the source end to adjust the
network’s traffic to a reasonable load level. A good combination of packet-drop
policies with traffic control mechanisms can maximize the throughput and
utilization of network and minimize the packet drop and delay.
Traditional packet-drop policy
The traditional packet-drop policy is tail-drop, that is, when the amount of packets
in a queue reaches the maximum value, all newly arrived packets are dropped.
This drop policy leads to global TCP synchronization, that is, when queues drop
packets of several TCP links at the same time, these TCP links enter congestion
avoidance and slow start status to adjust traffics simultaneously, and then reach
traffic peak simultaneously. In this way, network traffic keeps in frequent rises and
decreases.
RED and WRED
To avoid global TCP synchronization, random early detection (RED) or weighted
random early detection (WRED) can be used.
In RED algorithm, a maximum threshold and a minimum threshold are set for each
queue. The packets in the queue are processed as follows:
■ When the queue length is smaller than the minimum threshold, no packet is
dropped.