Network Support for High-Performance Distributed Machine Learning

The traditional approach to distributed machine learning is adapt algorithms the network, e.g., reducing updates curb overhead. Networks based on intelligent edge, instead, make it possible follow opposite approach, i.e., define logical network topology around task perform, so as meet desired performance. In this paper, we propose a system model that captures such aspects in context of supervised learning, accounting for both nodes (that perform computations) and information provide data). We then formulate problem selecting (i) which should cooperate complete task, (ii) number epochs run, order minimize cost while meeting target prediction error execution time. After proving important properties above problem, devise an algorithm, named DoubleClimb, can find $1+1/| \mathcal {I}|$ -competitive solution (with notation="LaTeX">$\mathcal {I}$ being set nodes), with cubic worst-case complexity. Our performance evaluation, leveraging real-world considering classification regression tasks, also shows DoubleClimb closely matches optimum, outperforming state-of-the-art alternatives.