摘要
随着大型电子业务系统和科研计算任务日益复杂,预测主机集群未来的指标数据变化趋势具有很强的现实意义.然而,主机时序数据的时空依赖问题很大程度上阻碍了主机集群资源的高效利用.为此,本文提出了TFSformer模型,该模型利用全局时空注意力和窗口卷积注意力解决了时空依赖问题.在时序数据预处理阶段,本文通过引入Wavelet-TCN-Embedding,实现了对主机负载特征序列的小波分解,将短期和全局时序依赖的特征有效地提取出来,从而提高了预测的准确性.其次,模型通过引入全局时空注意力和窗口卷积注意力,可以挖掘不同负载指标之间的空间依赖关系并且实现了对不同时间尺度下的时间依赖关系的关注,以此解决主机时间序列的时空依赖问题.同时,在主机系统上的实验证明了TFSformer在短期序列预测方面具备出色的性能,模型有效地提高了集群资源利用率并且降低了运维成本.
With the increasing complexity of large-scale electronic commerce systems and research computing tasks,predicting the future trends of indicator data for host clusters holds significant practical significance.However,the spatiotemporal dependency issue of host time series data largely impedes the efficient utilization of host cluster resources.To address this,this paper proposes the TFSformer model,which utilizes global spatiotemporal attention and windowed convolution attention to tackle the spatiotemporal dependency problem.In the preprocessing stage of time series data,this paper introduces Wavelet-TCN-Embedding to effectively extract features of short-term and global temporal dependencies from host load characteristic sequences,thereby improving prediction accuracy.Furthermore,the model introduces global spatiotemporal attention and windowed convolutional attention,which enable the exploration of spatial dependencies between different load indicators and the attention to temporal dependencies at different time scales,thus addressing the spatiotemporal dependency issue in host time series data.Moreover,empirical validation on host systems demonstrates the outstanding performance of TFSformer in short-term sequence prediction,effectively enhancing cluster resource utilization and reducing operational costs.
作者
赵卫东
潘智涛
张睿
吴乾奕
ZHAO Weidong;PAN Zhitao;ZHANG Rui;WU Qianyi(School of Software,Shanghai Key Laboratory of Data Science,Fudan University,Shanghai 200433,China)
出处
《小型微型计算机系统》
北大核心
2025年第6期1281-1288,共8页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(71971066)资助.
