Joint task planning and resource allocation in mobile edge learning systems

Abstract

Mobile Edge Learning (MEL) has recently emerged as a paradigm to enable distributed parallelized learning (PL) and federated learning (FL) on resource-constrained wireless edge devices. The development of distributed learning is a result of the fact that the number of devices connected to IP networks is increasing exponentially. Though the individual computing powers of these devices may be limited, their collective power is potentially unlimited. This unlimited yet under-utilized computing power coupled with MEL is the future technology for application and host serving. In this work, we aim to jointly optimize the planning of the learning process and tasks and the allocation of physical resources for mobile edge learning scenarios with global training time constraints. The term task planning refers to the number of local iterations in each of these global cycles and the number of data samples to be used for training by each device within each local iteration. The allocation of physical resources involves the determination of the computing speeds of each device and its communications resources to the MEL orchestrator. We discuss the problem of jointly optimizing the learning task planning and physical resources with two different objectives. In objective 1, we provide a solution to maximize the number of local training cycles each device executes within a given time constraint, which was shown to achieve a faster convergence to the desired learning accuracy. In objective 2, we provide a solution to minimize the global loss function of the training process by the end of the global training time constraint. Where we propose two novel algorithms, the dynamic and the static algorithm to solve the problem. We now show the performance of each objective compared to the results of optimizing the Task Planning (TP). Where we optimize the planning parameters and the allocation of tasks across the system and Equal Data Allocation (EDA), where we optimize the planning parameters and the physical resources allocation across the system.