Ph.D. Theses

End-to-End Multicast Congestion Control and Avoidance

By Jiang Li
Advisor: Shivkumar Kalyanaraman
July 3, 2003

IP multicast is an efficient mechanism to disseminate data to multiple recipients concurrently, where the source sends only one copy of the data and each receiver gets the data individually. One of the challenging research problems in IP multicast is to provide good and relatively simple congestion control/avoidance protocols, i.e. to keep the sending rate commensurate with the available bandwidth of the path between the source and each receiver. It is challenging due to the complicated patterns of congestion on different paths and the tremendous heterogeneity among them. Our research focuses on end-to-end solutions that do not require upgrade of routers.

We have proposed a series of multicast congestion control schemes for different situations. The first scheme LE-SBCC (Loss-Event oriented Source-based Multicast Congestion Control) is purely source-based and compatible with many multicast transport protocols. It is very easy to deploy because most of the functions are deployed on the source, while the minimum support of sending back packet loss indication is required on receiver side. The second scheme ORMCC requires more support from receiver side, but it avoids unnecessary feedback traffic from receivers to the source, and therefore can be used for larger multicast groups. The third scheme GMCC (Generic Multicast Congestion Control) extends ORMCC into a multi-rate framework where receivers can achieve different throughput. By leveraging ORMCC's control, the scheme has lower complexity than other proposed designs.

We have also studied multicast congestion avoidance where congestion is handled at its incipient stage before packet loss occurs (In congestion control, congestion is handled at its late stage when packets get lost). A scheme named MCA+ is provided as the solution, which is the first multicast congestion avoidance scheme to our knowledge.

To evaluate the schemes, we provide simulation results and analytical analysis. Some schemes have also been implemented and tested on Unix systems connected by real networks. In particular, we have run large scale simulations with up to 10,000 receiver nodes in a single multicast session, while the largest multicast simulations before only had one hundred receivers or so.

Return to main PhD Theses page