H3C S3610-28P Operation Manual page 1103

Operation Manual – QoS
H3C S3610&S5510 Series Ethernet Switches
Chapter 6 Congestion Avoidance
6.1 Overview
Serious congestion will bring great impact to the network resources, so some measures
must be taken to avoid congestion. As a type of flow control mechanism, congestion
avoidance monitors the utilization of network resources (such as queues or buffer in the
memory), and can drop packets when congestion deteriorates. In this way, the
congestion avoidance mechanism adjusts the network traffic so as to solve the
overloading problem in the network.
Compared to the port-to-port flow control, congestion avoidance controls more traffic
loading in the switch. When the switch drops packets from the source end, it can still
cooperate with the flow control actions (such as TCP flow control) on the source end so
as to adjust the traffic in the whole network to a reasonable load status. The
combination of packet drop policy and flow control mechanism on the source end can
maximize throughput and utilization rate of the network and minimize packet loss and
delay.
I. Traditional packet drop policy
Tail drop is adopted in the traditional packet drop policy. When a queue length reaches
the maximum value, all the new packets are dropped.
This packet drop policy will result in global TCP synchronization. If the queue drops
packets from multiple TCP connections simultaneously, these TCP connections will go
into the state of congestion avoidance and slow startup to reduce and adjust traffic and
then reach traffic peak in a certain future time. Such changes will cause the network
traffic jitter repeatedly.
II. RED and WRED
When congestion is too serious, the switch can adopt the random early detection (RED)
or weighted RED (WRED) algorithm to solve the problem of excessive congestion and
avoid global TCP synchronization caused by the tail-drop algorithm.
In the RED algorithm, an upper limit and a lower limit are set for each queue, and it is
stipulated that:
When the queue length is smaller than the lower limit, packets are not dropped.
When the queue length is bigger than the upper limit, all inbound packets all
dropped.
Chapter 6 Congestion Avoidance
6-1