Hybrid resource scheduling scheme for video surveillance in GPU-FPGA accelerated edge system

Abstract

In line with the development of ICT (Information Communication Technology), the core technology of the Fourth Industrial Revolution, the importance of smart cities is increasing to enhance public service and welfare. Among them, upgrading the intelligent surveillance system is considered an important task to enhance the safety and convenience of citizens. In particular, crowd-sourcing-based surveillance systems consist of a combination of complex tasks such as SSD, SIFT, and CNN character extraction, requiring a lot of computing workload. With these features, efforts have continued to accelerate the system. In traditional research systems, only cloud environments were considered, not edge environments with short communication delays, or heterogeneous accelerators were not considered to benefit from the service latency perspective. In this thesis, the technique of accelerating in edge computing system consisting of GPU and FPGA is studied for video surveillance system used in smart city. Considering the characteristics of the heterogeneous accelerator GPU and FPGA, each layer of CNN analyzes the benefits of resource management. A computation based execution time model for each CNN layer is established to predict the execution time for each CNN layer, and scheduling scheme is then presented to distribute this task graph to resources in consideration of resource availability. The algorithm aims to reduce the average execution time for a request and meet the execution deadlines. Based on the greedy algorithm, 1) gives each task a given deadline according to the maximum time path of the task graph, 2) divides tasks that can use the FPGA

distributes them to heterogeneous resources

and 3) executes run-time scheduling changes if problems occur, such as task execution time exceeds the given deadlines. The scheduling technique proposed in this thesis is experimentally compared with the previous GPU-FPGA scheduling scheme. With the previously proposed method, we compared the average running time efficiency, queue waiting time, resource utilization, and deadline violation rates. Experiments have shown that the proposed method is more efficient in average execution times than other models. Queue waiting time was improved up to 100 ms, and we were able to confirm that the deadline violation rate was lower than the prior model.