Tcp aimd algorithm (copy)

What is flow control )?

Traffic control focuses on how to dynamically adjust the sending rate of the source end to make full use of network resources without causing congestion. Therefore, throttling is a reasonable allocation of limited network resources among competing sources.

 

What is optimal flow control )?

Optimization Flow Control attempts to study and explain the flow control problem from the perspective of optimization theory. It regards the traffic control mechanism as a Distributed Algorithm for Solving Convex Programming Problems.

 

What are the key points of traffic control?

• Efficiency: traffic control algorithms should make full use of network resources;
• Fairness: users who use network resources should meet certain principles of fairness to minimize resource imbalance;
• Scalability: algorithms should be able to adapt to networks of different scales and will not be applied due to the increase in the network scale. This requires that the algorithm be distributed.
• Dynamic Performance: The traffic control algorithm will be applied in a complex time-varying environment, so the algorithm cannot be static and must have good dynamic performance. Generally, dynamic performance includes two aspects: stability and response speed.

 

What is TCP?

TCP is the traffic control mechanism of the Internet. The basic idea of the TCP protocol is that the source end can linearly increase the transmission rate to detect the remaining capacity of the network. At the same time, when congestion is detected, the transmission rate is reduced exponentially.

Taking TCP Tahoe as an example, the basic process is as follows: First, each connection is carefully sent from the window size of a single packet. Each time a confirmation packet is received, the window size increases by 1. In this way, the size of each RTT (round trip time) Time Window doubles. This is what we call TCP "slow start stage ". When the window size reaches a certain threshold, the source end enters the congestion avoidance phase. In this phase, the window size increases by 1 only when all the validation packets in the window are received. In this way, the size of each RTT time window is only increased by 1, which is the so-called additional growth. This determines when the threshold for "slow start" to enter "Congestion Avoidance" itself means the possible capacity of the Network, which will be adjusted accordingly based on the packet loss detected. Once packet loss is detected, the source end sets this threshold to half the current window size, re-transmits packet loss, and enters the "slow start" phase again.

 

What is AIMD principle?

The AIMD principle indicates that the TCP process will converge to a balance point that meets both efficiency and fairness.

Consider a system with only two users and one bottleneck. The system status can be expressed as follows. At the same time, any point indicates the resources allocated by the system to user 1 and user 2 (that is, the sending rate available to users ). When X1 + X2 = X, the system is most efficient. Corresponds to a line segment in the graph, which is called the efficiency line. When X1 = x2, the system has the highest fairness, corresponding to the fair line in the figure. The intersection of the fair line and the efficiency line is the control goal of the system. A straight line parallel to the efficiency line. The efficiency is a fixed value, called an equal efficiency line. The fairness of any line that goes through the origin is the same. It is called the equality line. The efficiency line is the overload area and the lower is the light load area.

Assume that the dual-user synchronizes the operation according to a certain cycle and follows the same rule. When packet loss occurs, the rate is reduced at the same time. Otherwise, the rate is increased at the same time. Consider the following methods:

Addition (AI): x (t + 1) = x (t) + A, a> 0
Addition subtraction (AD): x (t + 1) = x (t)-B, B <0
Multiplier increment (MI): x (t + 1) = C * x (t), c> 1
Multiplier subtraction (MD): x (t + 1) = D * x (t), d <1

We can see that the trajectory of AI and ad is a straight line of 45 degrees. AI increases fairness while moving to the overload zone. Ad reduces the fairness and moves to the Light Load area.

The trajectory of MI and MD is a straight line from the origin. Mi and MD are moving on fair online, so the fairness remains unchanged. Mi moves the track to the overload zone, while MD moves the track to the Light Load zone.

As shown in, the results of the AIMD process converge to the intersection of the efficiency line and the fair line. Therefore, TCP completes an efficient and fair resource allocation through addition, multiplication, and reduction.

 

Optimization of traffic control

Research Direction

-Joint optimization of traffic control and multi-path routing

Based on the optimized traffic control framework, Han studied the problem of multi-path routing. They proposed a source-side Algorithm for optimal Traffic segmentation between the source and the target.

-Joint optimization of traffic control and Physical Layer Power Control

Mung Chiang studies multi-hop self-organizing networks with limited link speed interference. He proposed a distributed power control algorithm, which improves the end-to-end throughput and network energy utilization through collaboration with TCP.

 

References

[Routing-Cross-layer] Han, H ., shakkottai, S ., hollot, C. V ., srikant R ., towsley, D.: overlay TCP for multi-path routing and congestion control. ENS-INRIA ARC-TCP workshop, Paris, France, (2003)

[Physical-Cross-layer] Mung Chiang. To layer or not to layer: balanling transport and physical layers in wireless multihop networks. Infocom 2004