Measuring the value of a research paper is usually difficult. Different
researchers might have different definitions of good research. However, no
matter how the definition varies from people to people, there is one judgement
that would be consistently applied -- research whose achievement impacts the
real world is definitely good research. Today, the congestion control protocol
proposed in this paper has been the default TCP algorithm for Linux and is
currently being used by more than 40% of Internet servers around the world and
by several tens millions Linux users for daily Internet communication. With no
doubt, this paper incurs a sound impact and is definitely classified as
excellent research under whatever definition.

Solid work does not need to be complex work. In fact, the proposed
congestion control protocol is very elegant, without any complex mathematical
analysis and theorems included. It uses a cubic function as the window growth
function to control the network congestions. The elegance of such design is to
achieve RTT-fairness (competing flows share the same elapsed time) and
intra-protocol fairness (two competing flows converge to a fair share).
Besides, the design offers TCP friendliness, backward compatible with
conventional TCP. No parameter adjustment are needed in practical engineering
implementations. These properties offer the great advantages in real systems:
a flow can be quickly recovered from the congestion, and the recovery is
scalable to networks with a large dynamic ranges of BDPs.

An interesting fact I noticed is that in the area of congestion control
protocol design, most influential works (e.g., STCP, HSTCP, HTCP, TCP-Vegas,
FAST, TCP-Westwood, TCP-Illinois, TCP-Hybla, TCP-Veno, as enumerated in this
paper) are of engineering style -- attacking a practical problem via intuitive
thinkings and designing instead of mathematical modelling and theoretical
deriving. Though I have limited knowledge in the spectrum of congestion
control, I can imagine that there must be tons of theoretical papers using
queuing theory, stochastic analysis and so on to model the design problem and
try to calculate to get the design. However, almost none of them finally
stands out to be influential to the real system. This poses a serious
question: does the theory really provide helpful guidelines in practical
design? My answer is that it depends. With no doubt, theoretical work could
deepen our understandings of real problem. But its utility is highly limited
by the impractical assumptions usually imposed, such as independent, identical
distributions, Poisson arrival rates, etc. In other words, to make the
analysis doable, it compromises accuracy, which could be an essential
problem in practice.