[NRG] practice talk, Mon, Oct 18, 4pm

vmanfred at bu.edu vmanfred at bu.edu
Thu Oct 14 10:09:16 EDT 2010


Folks,

I will be giving a practice talk (for T-Labs) on Monday, Oct. 18, at 
4pm in MCS 137. If you can come and give comments, that would be great. 
An abstract is below.

thanks,
Vicky

When a collection of  wireless or mobile nodes must form a network, a
number of different network-level approaches can be used to transport 
data from sender to receiver - stateful routing, an epidemic forwarding 
protocol such as flooding, or a delay tolerant network forwarding 
protocol such as store-carry-forward. We call the choice of which of 
these to use the network formation strategy. There is currently no 
precise understanding, however, of the specific network and traffic 
characteristics that determine (and a node might measure to determine)  
which network formation strategy is appropriate. For instance, how 
should nodes identify whether they are in a low or high mobility 
situation? Is average velocity always a sufficient measure, or is a 
measure based on how node velocity impacts link durations more 
informative?

In this talk, we  focus on the problem of  identifying  the empirical
measures that a node should estimate to determine which network 
formation strategy, either routing, flooding, or store-carry-forward, 
maximizes goodput. We argue that three empirical measures are 
sufficient to identify the appropriate network formation strategy: (1) 
the average number of senders, (2) the probability that an arbitrary 
route exists, and (3) the average link-up entropy. The average link-up  
entropy measures the value of routing information: it is the average 
conditional entropy of the current network state given the network 
state at some time in the past. We demonstrate on two different types 
of networks that these three measures reliably distinguish the 
situations in which each network formation strategy is typically 
appropriate, and that these metrics decompose the decision space in a 
consistent manner independent of network type.

This talk is based on joint work with Mark Crovella and Jim Kurose.





More information about the NRG-L mailing list