Elastic and Predictive Allocation of Computing Tasks in Energy Harvesting IoT Edge Networks

We consider a distributed IoT edge network whose end nodes generate computation jobs that can be processed locally or be offloaded, in full or in part, to other IoT nodes and/or edge servers having the necessary computation and energy resources. That is, jobs can either be partitioned and executed at multiple nodes (including the originating node) or be atomically executed at the designate server. IoT nodes and servers harvest ambient energy and jobs have a completion deadline. For this setup, we are concerned with the temporal allocation of jobs that maximizes the minimum level among all energy buffers in the network while meeting all the deadlines, i.e., that makes the network as much as possible energy neutral. Jobs continuously and asynchronously arrive at the IoT nodes, and computing resources are allocated dynamically at runtime, automatically adapting the processing load across nodes and servers. To achieve this, we present a Model Predictive Control based algorithm, where the job scheduler solves a sequence of low complexity convex problems and exploits future job and energy arrival estimates. The proposed technique is numerically evaluated, showing excellent adaptation capabilities, and performance close to that of an offline optimal scheduler with perfect information of all processes.