摘要
云计算环境下任务调度是优化云应用服务质量的热点研究问题,目前工业界和学术界重点关注任务调度策略。然而,现有方法依赖运维人员的系统实现知识或复杂的深度神经网络,需要较高计算资源,产生更高执行成本,难以适应动态变化的多样化任务类型。针对该问题,提出一种基于深度学习的云计算平台动态自适应任务调度策略。首先,从待处理任务、可用云资源及系统运行状态等三方面提取任务调度特征;其次,构建深度学习模型对特征编码,通过多头图注意力机制推理解码以预测策略的任务处理和调度执行成本;最后,根据调度收益从策略集中选择当前最优任务调度策略,同时基于迭代反馈机制计算损失函数以在线优化模型。建立虚拟化云计算服务器集群,实现典型的多种任务调度策略,模拟真实AI任务工作负载。实验结果表明,所提出策略与现有实验选取方法相比能够有效降低响应时间、执行成本及运行能耗。
Task scheduling in cloud computing environment is a hot research issue to optimize the quality of service of cloud applications.At present,industry and academia have proposed a variety of task scheduling strategies.Existing data-driven scheduling strategies rely on complex deep neural networks,which require high computing resources and incur higher execution costs,and are difficult to adapt to dynamically changing diverse task types.To solve this problem,we propose a dynamic adaptive task scheduling strategy for cloud computing platform based on deep learning.Firstly,the task scheduling features were extracted from three aspects of pending tasks,available cloud resources and system running status.Then,a deep learning model was constructed to encode the features,and the execution cost and response time of the strategy were predicted through inference and decoding.Finally,the current optimal task scheduling strategy was selected from the strategy set according to the scheduling profit,and the online optimization model of loss function was calculated.The experimental results show that the proposed strategy improves the execution cost,response time and energy consumption compared with the existing methods.
作者
任明
沈达
REN Ming;SHEN Da(China UnionPay Co.,Ltd.,Shanghai 201201,China)
出处
《计算机技术与发展》
2024年第8期17-22,共6页
Computer Technology and Development
基金
国家工业和信息化部重点专项(TC210804U-1)。
关键词
云计算
深度学习
任务调度
自适应策略
多头注意力
模型选择
cloud computing
deep learning
task scheduling
adaptive strategy
multi-head attention
model selection
