Multicast congestion control is a complex problem. However, a well designed multicast congestion control protocol is fundamental to the deployment of a multicast service. The major constraint while designing multicast congestion control protocols is the TCP-friendly constraint. The problem is that TCP-friendliness is not well adapted to large scale multicast sessions. When the protocol is source based, the TCP-friendliness is a real performance issue on large heterogeneous sessions. When the protocol is receiver based, TCP-friendliness is not well defined and lead to suboptimality.
We decided to explore alternatives to the TCP-friendly paradigm. Although this approach cannot offer a viable solution at short term, it offers an alternative to and a better understanding of the TCP-friendly paradigm. We introduced the notion of FS-paradigm. This paradigm relies heavily on previous works of Keshav and Shenker, but has a new focus on the general design of congestion control protocols. With the assumption of Fair Queuing in the network, we show that we can define nearly ideal end-to-end congestion control protocols. At first sight, the FS-paradigm statement seems pretty philosophical and difficult to apply. Indeed, it does not offer a practical solution to the problem of the design of congestion control protocols. However, it sheds a new light on the issues raised by TCP-friendliness and on alternatives to this TCP-friendly paradigm.
We applied the FS-paradigm to the design of a new multicast congestion control protocol called PLM. All the good properties predicted by the FS-paradigm are verified. PLM is based on the use of packet pair (PP) to infer the available bandwidth. We test PLM in a large variety of configuration and compare PLM with RLM and RLC. In all the cases studied, PLM outperforms RLM and RLC. The performance comparison was performed with both FIFO and FQ queues. Moreover, PLM is able to find the available bandwidth and to track the available bandwidth with no loss induced. To verify the stability of our PP bandwidth inference, we tested PLM in a complex environment (Self Similar and Multi-fractal background traffic). For all the scenarios considered PLM, never experienced any loss and followed closely the available bandwidth. It is clear that most of the good properties of PLM comes from the FQ network. However, even in a FQ network neither RLM nor RLC come close the PLM. Therefore this work has shown the importance of the bandwidth inference mechanism. The TCP-friendly paradigm imposes a bandwidth inference mechanism based on losses. We have shown that we can achieve a better efficiency with another class of bandwidth inference mechanisms.
In conclusion, the intent of this project was not to offer a practical alternative to the TCP-friendly paradigm, but to explore alternative design paths.
PLM has been in the ns release since July 24,
2000.
If you have a version of ns later than the July 24, 2000 daily snapshot,
PLM is already included into your ns version. If you have a version of
ns former than the July 24, 2000 daily snapshot, you need to follow
the procedure below in order to integrate PLM into your ns version.
This procedure guides you install PLM within your ns simulator. Please, follow the instructions below:
Warning: I only tested the installation of PLM on the versions 2.1b6 and higher of ns .
If you have any comments, questions, etc. I would be pleased to receive your message at: arnaud.legout@inria.fr