Reducing Energy Consumption of Wireless Sensor Networks through Processor Optimizations Gürhan Küçük and Can Başaran Department of Computer Engineering, Yeditepe University, 34755 Istanbul, Turkey {gkucuk, cbasaran}@cse.yeditepe.edu.tr Abstract- When the environmental conditions are stable, a typical Wireless Sensor Network (WSN) application may sense and process very similar or constant data values for long durations. This is a common behavior of WSN nodes that can be exploited for reducing their power consumption. This study combines two orthogonal techniques to reduce the energy dissipation of the processor component of the sensor nodes. First, we briefly discuss silent-store filtering MoteCache. Second, we utilize Content- Aware Data MAnagement (CADMA) on top of Mote- Cache architecture to achieve further energy savings and possible performance improvements. The com- plexity increase introduced by CADMA is also com- pensated by further complexity reduction in Mote- Cache architecture. Our optimal configuration reduc- es the total node energy, and hence increases the node lifetime, by 19.4% on the average across a wide varie- ty of simulated sensor benchmarks. Our complexity- aware configuration with a minimum MoteCache size with only four entries not only achieves energy sav- ings up to 16.2% but also performance improvements up to 14%, on the average. Index Terms- Computer Architecture, Sensor Networks I. INTRODUCTION Recent advances in process technologies and the shrinking sizes of radio communication devices and sen- sors allowed researchers to combine three operations (i.e. sensing, communication and computation) into tiny de- vices called wireless sensor nodes. Once these devices are scattered through the environment, they can easily construct data-oriented networks known as wireless sen- sor networks (WSNs). Today, there are a vast number of application scenarios involving WSNs in business, mili- tary, medical and science domains. The lifetime, scalability, response time and effective sampling frequency are among the most critical parame- ters of WSNs, and they are closely related to one crucial resource constraint that is very hard to satisfy: the power consumption. The WSN nodes are designed to be battery- operated, since they may be utilized in any kind of envi- ronment including thick forestry, volcanic mountains and oceanbeds. Consequently, everything must be designed to be power-aware in these networks. Small-scale operating systems, such as TinyOS [1], ambientRT [2], and computation-/communication- intensive applications significantly increase the energy consumption of the processor component of WSN nodes. Today, new sensor platforms with 16- [3] and 32-bit [4] processor architectures target more and more power- hungry applications. In [5], the researchers show that the processor itself dissipates 35% of the total energy budget of the MICA2 platform while running Surge, a TinyOS monitoring application. In [6], the authors claim similar energy values when running a TinyDB query reporting light and accelerometer readings once every minute. In [7], the researchers find that the energy consumption of the processor/memory component for raw data compres- sion is higher than the energy consumption of raw data transmission. Similarly, today most of the WSN applica- tions avoid extensive computations and choose to transfer raw data to server machines to increase the lifetime of the sensor nodes. On the contrary, our proposed design en- courages the WSN application developers to design less centralized applications by distributing the computation work and reducing the network traffic among the nodes. After observations of the results from our departmen- tal testbed and various simulations, we found two impor- tant characteristics of WSN data: 1. Temporal and value locality: Sensor network ap- plications have periodic behavior. Especially, monitoring applications, such as Surge, may sense and work on con- stant data and constant memory locations for long dura- tions. In this study, first we show that we considerably reduce the energy dissipation of the WSN processors by caching commonly used data in a small number of latches (MoteCache) [8]. Then, we also show that most of the store instructions in WSN applications are silent (these instructions write values that exactly match the values that are already stored at the memory address that is be- JOURNAL OF COMPUTERS, VOL. 2, NO. 5, JULY 2007 67 © 2007 ACADEMY PUBLISHER