GPU-Efficient Deployment of Ring All-Reduce-Based Distributed Model Training in Tidal Computing Power Network

This is the authors' abstract. We don't add key points for this paper.

This paper proposes a tidal-aware deployment algorithm for RAR-based DMT services in tidal CPN. The algorithm performance is evaluated in resource sufficient and constrained cases, respectively. Simulation results verify the benefit of reducing 20.6% GPU usage by dynamically partitioning training data and allocating GPU resources.