Abstract

Embedded systems measure noisy phenomena from the physical world and often generate outputs for control of noise-tolerant systems such as the human visual system. The algorithms that consume sensed data (e.g., pedometer algorithms) are often robust to limited input data errors. When the outputs of these algorithms are for human consumption, we can exploit both the robustness of algorithms to input errors as well as the flexibility of human perception, for more efficient sensor-driven interactive computing systems. And across both inputs and outputs, there is a missed opportunity to exploit the constraints imposed by the laws of nature and physical design of systems, to improve time-, energy-, and error-efficiency.

Lax, Rake, and Crayon are three recent systems that build on these observations to improve the energy efficiency of sensor activation, sensor data acquisition, and displays. Lax [SMR15a] reduces sensor operation power dissipation by over 40%, in exchange for infrequent sensor access failures that are easily masked by existing sensor applications. The Rake project [SMR15b, SMR16a, SMR16c]] trades data transfer power dissipation for data accuracy, minimally affecting the algorithms that consume this data. Crayon [SMR16b] reduces power dissipation of OLED displays, by exploiting the flexibility of human shape and color perception. Lax, Rake, and Crayon are a break from the conventional approach of focusing program optimizations in embedded systems on the programs themselves. By turning the focus of program transformations outward, to the hardware subsystems that provide program inputs (sensors) and the hardware that is the destination for program outputs (displays), our results show that it is viable and profitable to exploit the physics of signals in nature and the flexibility of human perception to make computing systems more efficient.

In this talk, I will overview the Lax, Rake, and Crayon systems and will briefly outline new ongoing work on a custom-designed hardware platform that builds on these ideas (Warp) and new description languages for describing the physical constrains of systems (Newton) and sensor access constraints (Sail).