From: Daniel Lowd (lowd@cs.washington.edu)
Date: Wed Oct 20 2004 - 01:02:46 PDT
This paper proposed an alternative to TCP that they called "XCP",
specially designed to cope with congestion on high bandwidth, high RTT
networks. The strength of their protocol is that they can achieve full
bandwidth utilization much more quickly by providing detailed congestion
feedback. The cost of this is more fields devoted to congestion control,
and intelligent routers that handle those fields. Their approach also
decouples efficiency and fairness: the router determines how much to
increase or decrease traffic (efficiency) and whose traffic to adjust
(fairness) separately.
It should come as no surprise that providing more information allows for
finer control. However, in the face of all the claims made by this paper
("It's always better in every scenario and lets you do more stuff!") I
found myself somewhat skeptical. For example, the stability analysis
assumed that every flow was identical: this is clearly unrealistic. The
experiments investigating a variety of different scenarios helped address
some of this in practice.
Furthermore, it seems that the routers are more vulnerable to malicious
hosts, since they depend on more information from them. A regular old IP
router doesn't have to believe any round trip times or window sizes. And
unless the XCP router keeps track of some per-flow state, it can never
tell if it's being fooled. The authors note that a router could
spot-check its connections, but a router can probably only afford that for
a small minority of its flows. Analogously, as long as highway patrolmen
are sufficiently infrequent, people will continue to speed.
Before deploying this widely, I would want to see a more detailed security
analysis -- what happens when, due to malice or negligence, a host lies
about its cwnd and rtt? What happens when many hosts lie, either
overestimating or underestimating? I noticed that the cwnd is never
smaller than a single packet. What happens if the host simply increases
their packet size to, say, 1 megabyte, and then fragments it? How hard is
it for a clever adversary to trick the fairness criteria into squeezing
out all other traffic?
The final challenge is, of course, deployment. I thought that aspect was
handled quite well. The authors described how "bilingual" routers could
offer XCP when the hosts were compatible and TCP congestion control
otherwise. This seemed fair.
Overall, I think this paper is useful for thinking about how networks
could be made more efficient by leveraging more information in routers.
The gains look impressive, and the experiments cover a variety of
scenarios, but I remain wary of what ill-performing hosts could do.
-- Daniel
This archive was generated by hypermail 2.1.6 : Wed Oct 20 2004 - 01:02:47 PDT