A Thread-Saving Schedule with Graph Analysis for Parallel Deep Learning Applications on Embedded Systems

The training of neural networks in embedded systems has become the direction of development nowadays. Because of the limited resources of embedded systems, it is not friendly to resource-consuming model training. What is more unfortunate is that during the training process, the allocation of thread resources by the operating system is not very appropriate, which can cause mis-allocation, and it will lead to wasting of resources or reducing performance during training. For the above problems, we propose a thread resource allocation method based on neural network graph analysis. Through the analysis of graphs, we dynamically allocate threads to achieve the full use of embedded system resources.