End-to-end Congestion Control for TCP-Friendly Flows with Variable Packet Size

Current TCP-friendly congestion control mechanisms adjust the packet rate in order to adapt to network conditions and obtain a throughput not exceeding that of a TCP connection operating under the same conditions. In an environment where the bottleneck resource is packet processing, this is the correct behavior. However, if the bottleneck resource is bandwidth, and flows may use packets of different sizes, resource sharing depends on packet size and is no longer fair. Now for some applications, such as Internet telephony, it is more natural to adjust the packet size, while keeping the packet rate as constant as possible. In this paper we study the impact of variations in packet size on equation-based congestion control and propose methods to remove the resulting throughput bias. We investigate in detail the design space of the approaches and propose a number of possible designs. We evaluate these designs through simulation and conclude with some concrete proposals. Our findings can be used to design a TCP-friendly congestion control mechanism for applications that adjust packet size rather than rate, or that are forced to use a small packet size.

End-to-end Congestion Control for TCP-Friendly Flows with Variable Packet Size (PDF).
Jörg Widmer, Catherine Boutremans, and Jean-Yves Le Boudec. ACM Computer Communication Review, April 2004

Related papers:

TCP Friendly Rate Control (TFRC) for Voice: VoIP Variant and Faster Restart , (draft-ietf-dccp-tfrc-voip-01.txt).
Sally Floyd and Eddie Kohler. Internet draft, work in progress, February 2005.

Variable Packet Size Congestion Control is included in the TFMCC code for the ns network simulator. It was implemented for version 2.26 but should work for 2.27/2.28 as well. For unicast simulations, simply use a TFMCC sender with a single receiver. Please read the instructions on how to add the code to an existing ns installation.

Usage:

VP-TFMCC has the following parameters:

VP-TFMCC uses either (1) a fixed packet size, or (2) a fixed packet rate, in which case the packet size varies (and the packet size parameter is unused). In this case, a maximum packet size can be specified as upper bound. For example, to make a TFMCC flow with packets of size 100 be fair to a TCP flow with packets of size 1000, set packetSize_ to 100.
```
Agent/TFMCC set packetSize_ 100
Agent/TFMCC set packetRate_ 0
Agent/TFMCC set packetMaxSize_ 0
      
```
To create a TFMCC flow with a fixed packet rate of 50 packets/second and a maximum packet size of 1000, set
```
Agent/TFMCC set packetSize_ 0
Agent/TFMCC set packetRate_ 50
Agent/TFMCC set packetMaxSize_ 1000
      
```
VP-TFMCC tries to achieve the same rate as a TCP flow sending packets of size fairSize_.
```
Agent/TFMCC set fairSize_ 1000
      
```
The following parameters determine how the receiver calculates a loss event rate from the received packets, as described in the CCR paper mentioned above. Note that there are different versions of the mechanisms, depending on whether the bottleneck is "packet mode" or "byte mode".
```
Agent/TFMCCSink set aggregatePackets_ false
Agent/TFMCCSink set aggregateLoss_ false
Agent/TFMCCSink set scaleLip_ false
Agent/TFMCCSink set randSampling_ false
Agent/TFMCCSink set unbiasInterval_ false
      
```
For "Unbiasing (Packet Mode)", set unbiasInterval_ to true. For "Virtual Packets (Byte Mode)" set aggregatePackets_ to true. For "Virtual Packets (Packet Mode)" set aggregatePackets_ and aggregateLoss_ to true. For "Random Sampling (Packet Mode)" set randSampling_ to true and for "LIP Scaling (Packet Mode)" set scaleLip_ to true.
The remaining congestion control mechanisms for byte mode have not been implemented.

All other parameters are as in normal TFRC or TFMCC.

Example scripts:

Some example code used for the CCR paper can be found here. Extract the archive, enter the directory, and start one of the simulations with ./run-scenario -sim fairness-aggloss-p100.