Congestion Avoidance is a QoS technique that aims to minimize network congestion (as the name implies) by proactively discarding certain packets used in TCP connections. To understand its functionality, it is essential first to know how TCP handles windowing and then explore how congestion avoidance mechanisms take advantage of this TCP process.

TCP Windowing

TCP windowing is a mechanism the protocol uses to manage the data flow between a sender and a receiver. It controls how much data the sender can send before waiting for an acknowledgment from the receiver. This process ensures that the sender does not overwhelm the receiver or the network with too much data at once.

As the sender transmits data as TCP segments, it waits for acknowledgments. Each acknowledgment indicates that the receiver has successfully received the data and is ready for more. Let's explain the process in a simplified way using the diagram below. 

Initially, the sender doesn't know the available network bandwidth up to the receiver or the receiver's capacity. That's why the sender first sends a tiny amount of data—typically one TCP segment (step 1). This process is referred to as the TCP Slow Start.

Every time the sender sends data and receives an acknowledgment, it tries to send more by doubling the number of segments. (steps 2 and 3). This process is referred to as Ramping Up. 

Eventually, this data transmission process reaches the point where the sender sends more data than the network can handle or the receiver can process. In both cases, the sender detects that the TCP segments were lost, as shown in the diagram below.

When TCP detects that a segment has been lost, the sender reduces the amount of data it sends by one-half. If multiple segments have been lost, the sender shrinks the sending amount by half numerous times. Then, the sender starts ramping up again by doubling the number of sent segments every time it receives successful acknowledgments.  

Keep in mind this is a simplified explanation of the process. It is much more complex at a low level and involves additional parameters. Also, there are multiple different implementations of the TCP protocol. However, the main point to remember is that a TCP data transmission between a sender and a receiver is not a straight line. It is a dynamic process of trying to utilize all the available bandwidth and the receiver's capacity, as shown in the diagram below.

It is especially important to understand the TCP windowing process at a high level because the QoS Congestion Avoidance takes advantage of this TCP process. 

Queue Drops

There is one more thing that we need to have as a context before jumping into the congestion avoidance techniques.

Queue drops occur when packets are discarded because a queue is unable to accommodate more packets. This happens when the queue reaches its maximum capacity. Let's discuss the different scenarios when a queue drops packets.

Tail Drops

Packets arrive at the tail of a queue. If the queue has space for new packets, they enter the queue and stay until they are scheduled for transmission. However, if the queue becomes full, new packets cannot be added and are dropped. This is called a "tail drop" because the packets arriving last, at the "tail" of the queue, are the ones that are dropped.

The diagram above illustrates the tail drop process. Notice that within an individual queue, the queuing mechanism is always FIFO - First In, First Out. If you don't follow the lessons in order and don't know what FIFO is, check out our previous lesson on Queueing and Scheduling. 

Head Drops

Once a packet enters a queue doesn't automatically mean it will be transmitted. A queue can also drop packets at the head. If the queue does not receive a scheduling slot for some time, packets age out and are dropped out of the queue from the head, as shown in the diagram below.

This can happen to a low-priority queue at times of congestion, depending on the scheduling mechanism in use. However, this dropping action doesn't typically occur with modern scheduling techniques such as LLQ and CBWFQ.

The queue can be drained entirely in extreme congestion situations to prevent traffic stalling and other resource overutilization problems. 

However, tail drops are by far the most commonly encountered queue drops that the interface experiences during congestion.

Why do we need Congestion Avoidance?

Knowing how TCP works at a high level and how a queue manages traffic in times of congestion is a prerequisite to understanding the QoS Congestion avoidance techniques.

The following diagram shows a graphical representation of the tail drop behavior of queues. The vertical axis represents the probability of a packet getting dropped. The horizontal axis shows the queue fill level in the context of arriving packets.

You can clearly see that the probability of a packet being dropped is zero percent until the queue fills up completely. Once the queue is full, the drop probability get 100% meaning newly arriving packets has 100% chance of getting dropped.

This behavior is pretty straightforward. There are two scenarios when this behavior is not desirable from a network point of view:

• When multiple TCP connections experience packet loss at the same time, their TCP window size shrinks to one-half at the same time. Then they start ramping up again. In the bigger scheme of things, this results in a period of massive underutilization of the network path, followed by a period of overutilization. The process typically repeats multiple times until the TCP connections are active.
• The second problem is that a queue typically holds multiple different types of traffic. When the queue drops traffic at the tail, it does not differentiate between the packets. It simply drops the newly arrived ones. This may impact business-critical applications, which is not desirable.

That's why a quality of service (QoS) tool called WRED has been introduced to prevent a queue from becoming completely full and starting to drop packets.

What is Weighted Random Early Detection (WRED)?

Weighted Random Early Detection, or WRED, is a method that prevents congestion and tail drops by taking advantage of the TCP windowing mechanism. It works by selectively dropping packets before the queue becomes full and experiences tail drops. The following diagram illustrates the WRED working logic in a graphical way. The horizontal axis represents the flow of arriving packets in the context of the queue fill level. The vertical axis represents the probability of a packet getting dropped.