MELOADES: Methodology for long-term online adaptation of embedded software for heterogeneous devices

摘要

In this work MELOADES [mel-uh-dees] is presented: a methodology for long-term online adaptation of embedded software that addresses the challenge of redeploying software and executing it within resource constraints. Instead of using fixed analytical models of resource consumption developed offline or tuning model parameters, MELOADES automatically reconfigures hardware online without any analytical model. MELOADES leverages long-term deployment by first selecting a set of hardware configurations that can potentially execute software tasks while satisfying a range of resource constraints and then storing these in a memoization table. The table is initialized using a Design of Experiments (DoE) survey to generate these speculative configurations. During deployment, for each new task assigned to the software, either a memoized configuration is found or a limited search for a new configuration that satisfies the task constraints is performed. Search results are added to the memoization table to reduce the time and energy required for future searches and eventually MELOADES converges to a simple table look-up. The effectiveness of this technique was demonstrated with an image capture and wireless transmission representative long-term application deployed on a Nokia N80 smartphone. Using a genetic search algorithm for energy efficiency/constrained image tasks, MELOADES satisfied 94% of all task constraints, evaluated only 1.6% of the configuration space, and used 98.5% less energy than an exhaustive search.