Date of Award

5-2016

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Computer Science

Committee Member

Dr. Brian Malloy, Committee Chair

Committee Member

Dr. Jason O. Hallstrom, Committee Co-Chair

Committee Member

Dr. Robert Geist

Committee Member

Dr. Jacob Sorber

Abstract

Wireless sensor networks (WSNs) typically consist of a large number of resource-constrained nodes. The limited computational resources afforded by these nodes present unique development challenges. In this dissertation, we consider three such challenges. The first challenge focuses on minimizing energy usage in WSNs through intelligent duty cycling. Limited energy resources dictate the design of many embedded applications, causing such systems to be composed of small, modular tasks, scheduled periodically. In this model, each embedded device wakes, executes a task-set, and returns to sleep. These systems spend most of their time in a state of deep sleep to minimize power consumption. We refer to these systems as almost-always-sleeping (AAS) systems. We describe a series of task schedulers for AAS systems designed to maximize sleep time. We consider four scheduler designs, model their performance, and present detailed performance analysis results under varying load conditions. The second challenge focuses on a fast and reliable network reprogramming solution for WSNs based on incremental code updates. We first present VSPIN, a framework for developing incremental code update mechanisms to support efficient reprogramming of WSNs. VSPIN provides a modular testing platform on the host system to "plug-in" and evaluate various incremental code update algorithms. The framework supports Avrdude, among the most popular Linux-based programming tools for AVR microcontrollers. Using VSPIN, we next present an incremental code update strategy to efficiently reprogram wireless sensor nodes. We adapt a linear space and quadratic time algorithm (Hirschberg's Algorithm) for computing maximal common subsequences to build an edit map specifying an edit sequence required to transform the code running in a sensor network to a new code image. We then present a heuristic-based optimization strategy for efficient edit script encoding to reduce the edit map size. Finally, we present experimental results exploring the reduction in data size that it enables. The approach achieves reductions of 99.987% for simple changes, and between 86.95% and 94.58% for more complex changes, compared to full image transmissions - leading to significantly lower energy costs for wireless sensor network reprogramming. The third challenge focuses on enabling fast and reliable data storage in wireless sensor systems. A file storage system that is fast, lightweight, and reliable across device failures is important to safeguard the data that these devices record. A fast and efficient file system enables sensed data to be sampled and stored quickly and batched for later transmission. A reliable file system allows seamless operation without disruptions due to hardware, software, or other unforeseen failures. While flash technology provides persistent storage by itself, it has limitations that prevent it from being used in mission-critical deployment scenarios. Hybrid memory models which utilize newer non-volatile memory technologies, such as ferroelectric RAM (FRAM), can mitigate the physical disadvantages of flash. In this vein, we present the design and implementation of LoggerFS, a fast, lightweight, and reliable file system for wireless sensor networks, which uses a hybrid memory design consisting of RAM, FRAM, and flash. LoggerFS is engineered to provide fast data storage, have a small memory footprint, and provide data reliability across system failures. LoggerFS adapts a log-structured file system approach, augmented with data persistence and reliability guarantees. A caching mechanism allows for flash wear-leveling and fast data buffering. We present a performance evaluation of LoggerFS using a prototypical in-situ sensing platform and demonstrate between 50% and 800% improvements for various workloads using the FRAM write-back cache over the implementation without the cache.

Share

COinS