From: Lillie Kittredge (kittredl@u.washington.edu)
Date: Tue Oct 26 2004 - 22:44:30 PDT
explicit allocation of best-effort packet delivery service
The authors describe their mechanism for prioritizing traffic, and using
those priorities to avoid congestion by preferentially dropping
low-priority traffic.
The authors have two priority levels: in and out. To decide which a
packet is, profile meters at the edge of the network observe how much
traffic has been sent and acked for a given flow, and if the sender is
over their "profile", (that is, if they're sending more vigorously than
they're allowed) the packets are marked as out and will be dropped
preferentially before the "legal" traffic.
The major benefit as I see it is a protection from bandwidth hogs.
Regular TCP is vulnerable to senders that fail to slow down when they get
the "slow down" signal (usually a dropped packet). With allocation
profiles, the percentage of "out" packets from a flow is a big help in
identifying misbehaving users, so they can be put in the doghouse and get
none of the bandwidth they're trying to gank, the louts.
It's a neat idea, and they show that it's got potential, especially in the
context of the RED gateways we talked about earlier, in which the in/out
scheme just increases the probability of extraneous packets getting
dropped before legit packets. I don't doubt that if it could get
deployed, it'd do well. But the question of whether it really can get
deployed is a fairly huge one. This requires changing routers, and adding
profile meters, though they may be implemented by changing existing
infrastructure. The other thing they require changing is frame formats,
to include the in/out bits.
I have the feeling that when it comes to congestion control, we're screwed
by the inertia of Internet infrastructure. This paper is yet another in a
line of ideas that would be great if they didn't need to change what we've
already got. Is there more work on congestion control in the context of
the existing infrastructure? If this paper had taken into account what a
partial deployment of their scheme would look like, I'd be appeased. What
to do when the packet lacks an in/out bit? What if this carefully
allocated traffic encounters a drop-tail router along the way? I'd like
to see these questions addressed.
This archive was generated by hypermail 2.1.6 : Tue Oct 26 2004 - 22:44:31 PDT