XCP Review

This paper proposes a new approach to fairness and congestion control
without explicitly designing for backwards compatibility or deployability in
the modern Internet. XCP is motivated by theory and experiments that show
that as bandwidth increases on the Internet and as higher delay networks are
connected (such as wireless networks), TCP doesn't reach full utilization
and becomes oscillatory and prone to instability, even with alternative
queuing schemes.

TCP can be inadequate for high bandwidth-delay product networks:

-- A flow's additive increase policy limits its ability to capture idle
bandwidth
-- It may take several RTT times to ramp up to full utilization, especially
grave for short flows
-- As users with different RTTs compete for the same bottleneck capacity,
the flows with higher RTTs will end up with lower throughputs.
-- Packet loss may indicate either congestion or error loss and a quick
response to a congestion loss cannot be made quickly

Finally, the aggressiveness of sources should be adjusted according to the
delay in the feedback loop. Any active queue management scheme with
constant parameters cannot be fast enough to operate on an arbitrary number
of TCP flows.

XCP attempts to solve these problems by separating fairness and congestion
control into two separate entities at the router. XCP routers precisely and
explicitly inform senders of congestion by modifying packet headers. The
receiver returns the updated congestion information to the sender in its ACK
packets. The efficiency controller only considers aggregate traffic over
the average RTT of the flows across a link and makes a single control
decision during a period equal to an average RTT. The fairness controller
is in charge of distributing the decisions among flows. In order to prevent
convergence stalling when efficiency is near optimal but distribution is
unfair, bandwidth shuffling is introduced.

XCP outperforms TCP in conventional and high bandwidth-delay environments.
It achieves fair bandwidth allocation, high utilization, small standing
queue size, and near-zero packet drops. It can be deployed incrementally in
the Internet using either clouds of XCP routers or by designing routers that
support both TCP and XCP. Additionally, because congestion is clearly
expressed in packet headers, the design of elements for punishing
misbehaving sources is simplified. In some situations, XCP may trade a few
percent of utilization in exchange for a much smaller queue size.
Received on Mon Oct 02 2006 - 19:56:33 EDT