GIT-CERCS-04-35
Matthew Wolenetz, Rajnish Kumar, Junsuk Shin, Umakishore Ramachandran,
Middleware Guidelines for Future Sensor Networks
In the near future, we envision sensor networks
to transport high bandwidth, low latency streaming
data from a variety of sources, such as cameras and
microphones. Sensor networks will be called upon to
perform sophisticated in-network processing such as image
fusion and object tracking. It is not too difficult to imagine
that computational capabilities of network nodes will scale
up relative to the fairly limited resources of current motes.
However, it is likely that energy will continue to remain a
constrained resource in such futuristic sensor networks.
Recently, there have been proposals for middleware
that provide capabilities for higher-level in-network processing
while minimizing energy drain on the network.
In this work, we analyze the interplay between resource
requirements for compute- and communication-intensive
in-network processing and resultant implications on figures
of merit of interest to an application including latency,
throughput, and lifetime. We use a surveillance application
workload along with middleware capabilities for data
fusion, role migration (simple relaying versus in-network
processing), and prefetching. Through a simulation-based
study, we shed light on the impact of device characteristics
such as CPU speed and radio features on application
figures of merit. We show, in the presence of prefetching,
that increasing radio bandwidth may not improve latency
nor throughput for compute-intensive workloads and may
actually decrease productivity of the network. We show that
cost function directed migration can significantly extend
application lifetime in sensor networks with topologies two
orders of magnitude larger than previous studies. We also
show that a simple minded cost function may not be
sufficient to guide migration decisions in the middleware.