End-to-End Congestion Control

Urs Hengartner

Diploma Thesis Summer 1997
Supervisors: J. Bolliger, Prof. T. Gross
Institute for Computer Systems, ETH Zürich

Objectives

In this diploma thesis, the congestion control algorithms of the transport protocols TLP and FACK TLP have been compared to several alternatives. TLP is a user-level protocol with congestion control mechanisms similar to TCP and FACK TLP additionally uses selective acknowledgments. Among these alternatives are:

redundant selective acknowledgments,
the continuous reduction of the congestion window during a recovery instead of an immediate decrease at its beginning (Rate-Halving),
a less aggressive data recovery algorithm,
a congestion avoidance strategy which reacts on the variation of the end-to-end delay caused by changes to the sending rate (TCP Vegas),
the setting of the first slow start threshold to the estimated bandwidth-delay product and
a burst avoidance using the spacing of back-to-back packets caused by the bottleneck router.

With tests running for two weeks and in which internet hosts distributed all over the world participated, some of these algorithms have been evaluated. Additionally, ack compression, i.e. the phenomenon of acknowledgments which arrive more closely spaced at the sender than the data packets they acknowledge at the receiver, has been examined.

Results

Redundant selective acknowlegment have been considered useful. This also applies to the new data recovery algorithm. The throughput of connections using Rate-Halving was smaller than the throughput of connections using the original algorithm. As a reason for this behavior, a weakness in the design of the (FACK) TLP sender has been identified. Connections which used TCP Vegas for congestion avoidance caused fewer retransmissions than connections with a steady increase of the congestion window. The mean round-trip time of TCP Vegas connections was smaller and less fluctuating, too. In contrary to other studies, in which the usage of TCP Vegas led to an increase of the throughput, in this evaluation the throughput even decreased. The estimation of the bandwidth-delay and its use as slow start threshold have been shown to be of small accuracy. Ack compression has been observed quite often. For the last three observations, it further has to be examined whether these results have been influenced by the design and the behavior of the protocol itself.

Author's address

Urs Hengartner
Sändli
CH-9657 Unterwasser

[ CS-Department | Up ]