Fully Distributed Deep Learning Inference on Resource-Constrained Edge Devices

摘要

Performing inference tasks of deep learning applications on IoT edge devices ensures privacy of input data and can result in shorter latency when compared to a cloud solution. As most edge devices are memory- and compute-constrained, they cannot store and execute a complete Deep Neural Network (DNN). One possible solution is to distribute the DNN across multiple edge devices. For a complete distribution, both fully-connected and feature- and weight-intensive convolutional layers need to be partitioned to reduce the amount of computation and data on each resource-constrained edge device. At the same time, resulting communication overheads need to be considered. Existing work on distributed DNN execution can not support all types of networks and layers or does not account for layer fusion opportunities to reduce communication. In this paper, we jointly optimize memory, computation and communication demands for distributed execution of complete neural networks covering all layers. This is achieved through techniques that combine both feature and weight partitioning with a communication-aware layer fusion approach to enable holistic optimization across layers. For a given number of edge devices, the schemes are applied jointly such that the amount of data to be exchanged between devices is minimized to optimize run time. Experimental results for a simulation of six edge devices on 100 Mbit connections running the YOLOv2 DNN model show that the schemes evenly balance the memory footprint between devices. The integration of layer fusion additionally leads to a reduction of communication demands by 14.8%. This results in run time speed-up of the inference task by 1.15x compared to partitioning without fusing.