From: Ethan Phelps-Goodman (ethanpg@mac.com)
Date: Sun Nov 14 2004 - 17:44:56 PST
Internet Indirection Infrastructure
Stoica et. al.
This paper presents an general solution to the problems of multicast, anycast, server selection, and mobility. These seemingly separate problems are all instances of indirection. Rather than work at the IP level, where it may be nearly impossible to deploy a new system, the authors advocate a general-purpose overlay network that could handle indirection in a variety of settings. The particular model of indirection is simple: senders send to abstract IDs, and receivers register to receive packets sent to a particular ID. (There are more complicated mechanisms involving inexact matching and multiple IDs, but the basic idea is the same.) This is implemented as a distributed hash table, which provides scalable distributed key-value lookup.
The protocol has two huge strong points. First, it is almost trivial to deploy. Second, it is extremely general. Some of the less obvious examples given of possible uses are providing transparent intermediary servers for uses such as WML conversion and video rate conversion and providing transparent server selection based on geographic location. For its primary uses however, multicast and mobility, their arguments seem weak. There is no discussion of how an efficient multicast tree would be constructed using i3. There is some talk of load balancing, but this seems a far cry from attempting to build an efficient distribution tree. (They do cite another paper of theirs specifically on multicast with i3, so presumably more work is done there.) For mobility it's not clear to me that they solve a pressing problem--it is easy enough to have a fixed proxy server forward to a mobile node. There is also the problem that all flows now have to be routed through an intermediary node. With some probing of the network fo
r low latency hosts, they can reduce the latency increase to a factor of 2, but I'm vaguely uncomfortable with their mechanism for doing this. (Chord assumes that keys are distributed uniformly over the namespace, and probing for latency could dramatically alter this assumption.) Finally, it seems like there would be a big practical problem related to the heterogeneity of hosts. Even with replication of keys, imagine the poor cable modem user who happens to get saddled with CNNs video multicast.
Ethan
This archive was generated by hypermail 2.1.6 : Sun Nov 14 2004 - 17:45:02 PST