SenseBench: toward an accurate evaluation of sensor network processors

Sensor network processors introduce an unprecedented level of compact and portable computing. These small processing systems reside in the environment which they monitor, combining sensing, computation, storage, communication, and power supplies into small form factors. Sensor processors have a wide variety of applications in medical monitoring, environmental sensing, industrial inspection, and military surveillance. Despite efforts to design suitable processors for these systems (Ekanayake et al., 2004; Hempstead et al., 2005; Nazhandali et al., 2005; Wameke and Pister, 2004), there is no well-defined method to evaluate their performance and energy consumption. The historically used MIPS (millions of instructions per second) and EPI (energy per instruction) metrics cannot provide an accurate comparison because of their dependence on the nature of instructions, which differ across instruction set architectures. On the other hand, the current well-defined benchmarks (1989; Guthaus et al., 2001; Lee et al., 1997) do not represent typical workloads of sensor network systems, and hence, are not suitable to compare sensor processors. This paper defines a set of stream applications representing the typical real-time workload of a sensor processor. Furthermore, three new metrics, EPB (energy per bundle), xRT (times real-time), and CFP (composition foot print) are introduced to evaluate and compare such systems.