Uncertainty-Aware Multitask Allocation for Parallelized Mobile Edge Learning

Harvesting the profuse yet underutilized computational resources of IoT devices, also referred to as Extreme Edge Devices (EEDs), can significantly curtail the delay in parallelized Mobile Edge Learning (MEL). However, EEDs are user-owned devices, which causes them to experience a highly dynamic user access behavior. Such dynamicity can lead to uncertainty in the available computation and communication capabilities of learners. In this paper, we propose the Minimum Expected Delay (MED) scheme. MED is the first data allocation scheme in MEL that accounts for uncertainty in learners' capabilities and enables multitask allocation. Given the state probabilities of learners, MED strives to minimize the sum of the maximum expected delay of all tasks, while abiding by certain training time and budget constraints. Towards that end, MED formulates the data allocation problem as an Integer Linear Program (ILP) and makes uncertainty-aware decisions. We conduct rigorous experiments on a real testbed of Jetson Nano devices. Extensive performance evaluations show that MED outperforms a representative of state-of-the-art uncertainty-naive schemes by up to 11%, 11%, 42%, and 5% in terms of training time, satisfaction ratio, data drop rate, and occupancy time, respectively. In addition, MED approaches a baseline scheme that assumes a perfect knowledge of the learners' states, yielding a gap of up to 10%, 5%, and 14% in terms of satisfaction ratio, data drop rate, and occupancy time, respectively.