本文已被:浏览 139次 下载 928次
投稿时间:2023-03-03 修订日期:2023-07-29
投稿时间:2023-03-03 修订日期:2023-07-29
中文摘要: 本文考虑当前研究都忽略的边缘节点计算资源对任务分析延迟、准确度相悖的影响所带来的权衡问题,提出一种动态配置视频分析任务在边缘的带宽和计算资源的分配策略,通过资源在线分配,实现延迟与准确度的最佳权衡。方案通过优化带宽与计算资源目标,最大化目标函数,即在降低延迟的情况下最大化准确率。其中由于显示数据集缺乏,利用模拟、仿真技术,通过最小化误差函数法获得了准确度函数拟合;同时利用梯度估计法,找到目标函数的梯度下降方向,通过不断迭代来求解最小值,以解决神经网络不可知性带来的无法直接获得梯度的问题。最终通过仿真对照试验验证了算法的优越性,能够高效降低网络负荷,提高整体资源利用率及性能水平。本文同时讨论了以边缘-云协作的架构提高资源利用率,并提出了未来研究方向。
Abstract:The paper considers the trade-off between the impact of edge node computing resources on task analysis latency and accuracy, which is currently overlooked in research. Hence, by proposing a dynamic configuration strategy for allocating computing resources for video analysis tasks at edge nodes, the article tries to realize the best balance between tasks latency and accuracy through online resource allocation. The solution maximizes the objective function by optimizing computing resource objectives, i.e. maximizing accuracy while reducing latency. Due to the lack of data sets, the Error function is minimized by utilizing the simulation to fit the accuracy function. At the same time, the gradient estimation method is used to find the gradient descent direction of the objective function, and the minimum value is solved through iterations to solve the problem of not being able to directly obtain the gradient caused by the unpredictability of the neural network. The superiority of the algorithm is verified through simulations and control experiments, which can efficiently reduce network load and improve overall resource utilization and performance level. The paper also discusses situations of improving resource utilization with the edge-cloud collaboration architecture, and investigates the future research direction.
文章编号: 中图分类号: 文献标志码:
基金项目:
作者 | 单位 | |
孙越* | 中国华电科工集团有限公司 | suny@chec.com.cn |
何牧 | 中国华电科工集团有限公司 | |
庞琦方 | 中国华电科工集团有限公司 |
Author Name | Affiliation | |
Sun Yue | China Huadian Engineering Co., LTD. | suny@chec.com.cn |
He Mu | China Huadian Engineering Co., LTD. | |
Pang Qifang | China Huadian Engineering Co., LTD. |
引用文本: