HeterPS: Distributed deep learning with reinforcement learning based scheduling in heterogeneous environments

Deep neural networks (DNNs) exploit many layers and a large number of parameters to achieve excellent performance. The training process DNN models generally handles large-scale input data with sparse features, which incurs high Input/Output (IO) cost, while some are compute-intensive. exploits distributed computing resources reduce time. While heterogeneous resources, e.g., CPUs, GPUs multiple types, available for the process, scheduling diverse remains critical process. To efficiently train model using we propose framework, i.e., Heterogeneous Parameter Server (HeterPS), composed architecture Reinforcement Learning (RL)-based method. advantages HeterPS three-fold compared existing frameworks. First, enables efficient workloads resources. Second, an RL-based method schedule workload each layer appropriate minimize cost satisfying throughput constraints. Third, manages storage communication among We carry out extensive experiments show that significantly outperforms state-of-the-art approaches in terms (14.5 times higher) monetary (312.3% smaller).