[SensorNetArch] On the Common Case and Making it Fast

[Cheng Tien Ee]

Hi,

I've been thinking about the services the architecture should provide to 
applications, and I think it should depend on the type of data that's 
carried by the network. For instance in the Internet we have TCP and 
UDP. What I'm going to say next is definitely still being thought 
through, but mayhaps there will be useful feedback.

Traffic can be broadly classified into 2 categories: low latency and low 
reliability, and high latency (or latency can be ignored) and high 
reliability. The definition of latency seems straightforward but not 
reliability. Reliability can be a combination of link-level and 
end-to-end. Currently I don't believe anyone has done end-to-end 
reliability, or perhaps there's no (and never will be the) need to. It 
is also clear that end-to-end without link-level has great potential in 
causing terrible performance in a wireless sensor network. Most traffic 
in a sensor network should require low latency and low reliability: 
small amounts of data (such as sensor readings) are gathered at either 
the base station or an intermediate mote. We don't need all packets ever 
generated to be delivered, but we want enough of them. Since the 
assumption is that small amounts of data are sent at a time, they are 
likely to fit in small packets of size that of the link-level MTU. And 
since this is likely to be the case in general, I would argue that the 
architecture's basic service (let's call this BS :) would be to provide 
transmission of small payloads (that can fit in a link-level packet). 
This service should be directly visible to the application layer, i.e. 
we don't want to force the application to use another service that then 
uses BS.

The architecture can then provide additional services (layers?) above 
this BS. For instance, if the application aggregating data and gathering 
statistics from thousands of motes decides that its one-per-24-hours 
payload is long and require it to be more reliabily transported to the 
base station then an additional component/layer can be built on BS and 
used to provide fragmentation and FEC. The overhead of doing this is not 
great considering all the traffic carried because this is not the common 
case.


Ee