| Title |
Performance Improvement of Multiple DNN Models inside Embedded Systems through CPU-GPU Collaborative Scheduling |
| DOI |
https://doi.org/10.5573/ieie.2021.58.1.33 |
| Keywords |
Deep neural network; CPU-GPU; Collaborative scheduling; Embedded system |
| Abstract |
The use of deep neural network (DNN) models has improved the performance of many applications. To get the better performance, research on the DNN model itself is actively progressing, but research on a system that can perform this efficiently is relatively insufficient. In order to satisfy the high recognition success rate and the requirements of various applications, various types of DNN models are used inside embedded systems such as robotics and autonomous vehicles, and the number itself is increasing. Therefore, when a DNN accelerator such as a GPU is shared and used by multiple applications, a performance bottleneck due to the GPU is bound to appear. In this paper, to solve this problem, we propose a scheduling framework that can utilize both CPU and GPU for DNN model computation. This technique uses a method of minimizing data transmission overhead between cores of different types by selecting a core type suitable for the computational characteristics of each layer of DNN models. As a result of experimenting with the proposed technique on an actual commercial board, it is up to 71.1% higher than before applied. |