TCP Nice

Paper Review: TCP Nice: A Mechanism for Background Transfers

This paper describes a new system, TCP Nice, which allows for large
background transfers used to improve service quality. Applications
can trade bandwidth consumption for improved service latency,
availability, consistency and mobility support. Operating systems do
not provide support for aggressive background transfers since they can
interfere with foreground transfers. A delicate balance must be found
between foreground and background transfers for optimal performance
and to avoid self-interference and cross-interference.

TCP Nice, like Vegas, attempts to detect congestion before reaching
packet loss. Like Vegas, it infers congestion conditions from
increasing RTT values; however, the paper argues that Vegas? scheme
doesn?t prevent cross-interference. It addresses this issue with
three measures. It uses a more sensitive congestion detector,
multiplicative decrease when RTTs increase and allows for congestion
windows smaller than one.

TCP Nice is essentially a conservative extension of Vegas. Although
it is capable of supporting aggressive background flows, it may result
in underutilization of bandwidth. Also, in the face of packet loss
not due to congestion, TCP Nice?s actions are perhaps too extreme and
will undoubtedly reduce throughput.

Simulations are conducted to further the support for TCP Nice. The
first simulation shows that in terms of latency, TCP Nice approaches
Router Prioritization (ideal) in the case of utilizing spare
bandwidth. When spare bandwidth was fixed and the number of
background flows was varied, the background flows had a negligible
effect on foreground flow traffic for Nice while significantly
affecting Reno and Vegas foreground traffic. One caveat in the
simulations was the fact that when traffic was unpredictable, the
benefits of Nice were reduced. Traffic in the Internet is considered
highly unpredictable and thus real-world application of TCP Nice may
not result in significantly improved performance. In any study
relating to the Internet, it is always preferential to have real world
experiments rather than to rely on simulations which make use of
models and assumptions.

Two case studies were examined to show the benefits of TCP Reno under
real-world applications rather than simulated models.

Another advantage of the work is that it takes an end-to-end approach
for accommodating background transfers rather than requiring
router-support which is unrealistic and unlikely to ever be adopted.
Although support for background transfers is an attractive feature,
creating a less aggressive version of TCP may not be the best approach
as it is unclear just exactly how much of a factor congestion in the
Internet is. The fact that TCP Nice can coincide with other
implementations of the protocol (i.e. Vegas, Reno) makes it more
attractive and gives it a significant chance to be utilized in
machines running background transfers regularly.

-- Nadeem Abji
Received on Wed Oct 04 2006 - 23:04:59 EDT