Clank: Architectural support for intermittent computation

The processors that drive embedded systems are getting smaller; meanwhile, the batteries used to provide power to those systems have stagnated. If we are to realize the dream of ubiquitous computing promised by the Internet of Things, processors must shed large, heavy, expensive, and high maintenance batteries and, instead, harvest energy from their environment. One challenge with this transition is that harvested energy is insufficient for continuous operation. Unfortunately, existing programs fail miserably when executed intermittently. This paper presents Clank: lightweight architectural support for correct and efficient execution of long-running applications on harvested energy-without programmer intervention or extreme hardware modifications. Clank is a set of hardware buffers and memory-access monitors that dynamically maintain idempotency. Essentially, Clank dynamically decomposes program execution into a stream of restartable sub-executions connected via lightweight checkpoints. To validate Clank's ability to correctly stretch program execution across frequent, random power cycles, and to explore the associated hardware and software overheads, we implement Clank in Verilog, formally verify it, and then add it to an ARM Cortex M0+ processor which we use to run a set of 23 embedded systems benchmarks. Experiments show run-time overheads as low as 2.5%, with run-time overheads of 6% for a version of Clank that adds 1.7% hardware. Clank minimizes checkpoints so much that re-execution time becomes the dominate contributor to run-time overhead.