Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

   

2009 Research Summary

System Architecture Directions for Storage-Centric Sensor Networks

View Current Project Information

Prabal Kumar Dutta, David E. Culler, Joseph M. Hellerstein, Scott Shenker and Ion Stoica

National Science Foundation

Many sensor networks deployed to date stream data from sources to sinks. This work argues that the streaming-centric paradigm is not well suited to many energy-constrained, yield-critical, high-throughput, high-concurrency, or occasionally-disconnected systems and that a storage-centric architecture would better serve delay-insensitive applications without sacrificing the performance of real-time ones. Our hypothesis is that placing storage elements like sample-and-holds, queues, and buffers along sensing, query processing, and network data paths will significantly improve system performance for typical workloads and enable new applications not practical using current architectures. Storage facilitates temporal indirection or delay, allowing us to break the synchrony of streaming systems. Artificial delay often does not harm the end user and provides optimization opportunities. If data are batched, its size can be reduced through compression, which reduces intermediate buffering requirements; its unit of reliable transfer can be bundles rather than packets, which better amortizes channel acquisition cost; and its transport can occur over links whose quality does not change appreciably during the transfer, which improves reliability.

The availability of cheap, efficient, and plentiful storage in the form of flash memory behooves us to reconsider the design space for sensor networks beyond basic efficiency, throughput, and reliability considerations. Sensing and storage bandwidth already exceed the end-to-end bandwidth available for streaming systems that employ multihop wireless communications, and this gap is only expected to increase. At the same time, new applications that require high sampling rates, mobile or disconnected operation, opportunistic communication, shared sensing infrastructure, and ad hoc coordination are emerging. Storage will play a central role in these new applications as sensor nodes become the primary replica of data for want of bandwidth and an embarrassment of storage riches--the spot price for 8 Gb (1 GB) of NAND flash on May 15, 2007 was $7, compared with $22 and $56 spot prices one and two years ago, respectively. Going forward, this trend suggests that we should think of sensors as not just sources of streaming data but also as potentially deep reservoirs of historical data to be mined without the need to collect it all centrally.

[1]
P. Dutta, D. Culler, and S. Shenker, "Procrastination Might Lead to a Longer and More Useful Life," HotNets-VI, Atlanta, GA, November 14-15, 2007.