Yield over Energy: Task-based Intermittent Computing with Themis

We present Themis: a task-based intermittent programming system able to steer task executions based on programmer-provided directives on the tasks’ relative data yield. Unlike existing solutions that solely seek energy conservation, Themis allows programmers to slice the application logic into task groups with attached data yield directives. Themis takes these information as an input and regulates the execution of task groups against energy failures and periods of energy surplus. We compare Themis against state-fthe-art task-based intermittent programming systems using five real-world energy traces, measuring how closely each system can match the programmer directives.We showthat Themis is up to 31% closer to the programmer directives compared to the best performing baselines and, depending on available energy, can at times even match those requirements perfectly. This performance is achieved while responding to wild energy fluctuations and without requiring manual fine-tuning on a per-scenario basis.