(no subject)

This paper outlines a protocol for the congestion-controlled sending
of unreliable data named DCCP. The protocol seems intended to
partially replace UDP unicast, by providing a better protocol for
persistent communication. It is suggested that the need for such a
protocol has arisen because there were few if any applications
requiring persistent unreliable connections until quite recently,
when streaming media and communications started becoming prevalent on
the web.

The authors began with the assumption that designing an unreliable
congestion-controlled protocol could be done by stripping the
reliability from TCP. They discovered, however, that TCP reliability
and congestion control were closely linked and were forced to largely
start over.

Perhaps the most significant change as compared to both UDP and TCP
is the acknowledgement of received packets. Whereas UDP includes no
acknowledgement mechanism and TCP acknowledges packets by specifying
the lowest-numbered packet that has not been received, DCCP
acknowledges the latest received packet, which has significant
repercussions in the overall design of the protocol. In general,
though, the protocol seems much closer to TCP in that it is a
connection-oriented protocol.

One of the neat features of DCCP is the modularity of its congestion
control mechanisms. The protocol was designed to allow applications
to negotiate their congestion control mechanisms and even use
different mechanisms in each direction. The two mechanisms currently
provided are a TCP-like mechanism designed for applications like
games for which it is most important to use all of their available
bandwidth, and a smoother mechanism designed for applications such as
voice-communication where sudden reductions would be harmful,
although they suggest a range of application requirements for which
other mechanisms might be developed.

The weakness of this paper is that it lacks quantitative results. It
would be valuable to see how the bandwidth, delay, and computational
overhead at the endpoints of DCCP compares to that of TCP and UDP
over mixed networks of various proportions. I'd be curious to know
whether the acknowledgements for one-way communication significantly
affect network congestion.
Received on Mon Oct 09 2006 - 17:37:26 EDT