From: Chandrika Jayant (cjayant@cs.washington.edu)
Date: Tue Oct 19 2004 - 23:53:11 PDT
"Congestion Control for High Bandwidth-Delay Product Networks"
Written by Dina Katabi, Mark Handley, and Charlie Rohrs
Reviewed by Chandrika Jayant
This paper introduces a solid step up from TCP congestion
control with the creation of XCP, a stable, adaptive, flexible,
efficient, low-drop, high performance, and low-queuing protocol which
can stand up to large bandwidth and delay increases. It also claims to
improve performance in conventional environments where TCP does pretty
well. TCP's congestion control, for all practical purposes, collapses
when faced with these more demanding networks. The beauty of XCP, other
than its amazing results compared to TCP, is that is it simple to
implement, flexible working in dynamic environments, and can work as its
own pure network or coexist with existing TCP framework, allowing for
theoretical graceful evolution (though routers would need added support,
which is not a simple task..). To provide flexibility, fairness and
utilization are decoupled- a crux of this paper. Instead of pushing a
network to its limit (unintuitive approach), there is feedback in packet
headers- degrees of congestion, not Boolean values, are measured- and
packet drops due to congestion are very rare.
My immediate attraction to this paper was that it was
centered around control theory rather than handwaving or fudging. The
system converges to stability independent of delay, bottleneck capacity,
and number of sources, allowed by its constant parameters alpha and beta
(and is still robust with regard to estimation errors). This seems
intuitively elegant to me and even if there are obstacles in deploying
this system, the ideas presented are a good model of how systems in the
future should be thought out.
I wish the authors had not just settled for 2 specific
control laws for EC and FC just because they worked. At least I wish
they had mentioned some others that they think would be useful to try
out. This could be some future work. I would also like to see different
allocation schemes experimented with in this context. It is interesting
to note that security is briefly touched upon, a "newer" concern in
networks as compared to the papers we've been reading in class. However,
it is just glazed over and I'd like to hear more about it.
This is a very important paper because network congestion
has been a problem in the forefront for almost twenty years, and as
networks keep expanding so does the level of service we need for them.
This paper introduces a protocol getting closer to the ideal for service
differentiation, a very pressing need. XCP violates the pure end-to-end
argument, so it will be interesting to see how that plays out in the
network community.
This archive was generated by hypermail 2.1.6 : Tue Oct 19 2004 - 23:53:19 PDT