Keynote 2

In principle, the entire world can earnestly exploit sensors and pervasive systems to great societal benefits. In practice, however, there are no widely known models or tools for evaluating, simulating or benchmarking their performance. Unlike traditional distributed systems, where the resource/needs of computation and communication dominate the performance equation, pervasive systems forge new metrics and requirements. First, sensing should be accounted for, in addition to computing and communication. This includes sensing latency and sensing energy consumption. Second, sentience efficiency should be established as a critical metric of performance; oversampling a sensor or unnecessarily sensing another decreases both sentience efficiency and energy efficiency. In this keynote, I will present a performance model of pervasive systems based on a symbiotic relationship between the massive-scale networks of sensors and the cloud. I will show why initial processing of sensor data must take place deep in-network, or the cloud could be bottlenecked. I will also show the role of edge computing and how it connects sensors to the cloud, enables pervasive decisioning and improves performance. I will present the Atlas sensor platform, an in-network sensing, communication and processing technology utilized to implement the cloud-edge-sensor collaboration model. I will then report on several performance evaluation studies based on Atlas, including in-network sensor query optimization, in-network phenomena detection and tracking and in-network event-condition-action rule evaluations. Finally, I will present Persim, a recent effort to create pervasive system simulation environments with focus on activity recognition and dataset generation and validation. I will end with a summary of grand challenges and opportunities in the area of pervasive systems performance modeling and evaluation.