| Abstract: | High-power sensors, such as steerable radars and pan-tilt-zoom cameras, expose programmable actuators to applications, which actively control them to dictate the type, quality, and quantity of data collected. Since networks of high-power sensors are expensive to construct, maintain, and deploy, they represent a scarce resource that multiple applications with conflicting goals often share. However, existing mechanisms are too coarse-grained to satisfy these conflicting goals. To address the problem, this paper presents VSense, a virtualization approach that augments virtual machines with virtual sensors to enable fine-grained multiplexing between unmodified applications. VSense represents each virtual sensor as a state machine, interleaves state transitions (i.e., acutations) to balance fair access to the physical sensor with its efficient use, and employs an actuation-aware proportional-share scheduler to ensure performance isolation between virtual sensors while optimizing actuation overheads. We implement VSense in Xen and prototype it using one example of a high-power sensor with actuators---a PTZ camera---to multiplex concurrent applications. Our results show that VSense efficiently isolates the performance of virtual sensors, allowing concurrent applications to satisfy conflicting goals. In a case study we conduct, we show that concurrent applications receive tolerable, but degraded performance---ranging from 1.5x less to 8x less in our examples---than seen with a dedicated sensor.
|
