Decentralized Data Allocation via Local Benchmarking for Parallelized Mobile Edge Learning

Multi-Access Edge Computing (MEC) has emerged as a computing paradigm that can facilitate the use of Mobile Edge Learning (MEL), where Machine Learning (ML) models are processed at the edge. In MEL, it is important to address system heterogeneity in a way that minimizes staleness to improve learning accuracy. To do so, a centralized data allocation approach is typically used. However, this approach tends to overlook the privacy of learners, since learners' capabilities are assumed to be known beforehand by the orchestrator. In this context, we propose the Data Allocation via Benchmarking (DAB) scheme. DAB is a decentralized data allocation scheme that eliminates staleness and achieves a certain QoS while preserving the privacy of learners. DAB does not allow any information about the learners to be known to the orchestrator. Instead, each learner estimates the upper bound on the amount of data that it can train such that a certain training deadline is not exceeded. In addition, DAB proposes a novel method to enable each learner to accurately estimate its own hardware characteristics via benchmarking. Extensive performance evaluations on a real testing environment have shown that DAB can outperform the centralized data allocation scheme by up to 12% and 26% in terms of loss and prediction accuracy, respectively. Performance evaluations also show that the proposed benchmarking scheme yields an 83% reduction in benchmarking error compared to a prominent baseline scheme.