摘要
微服务的多实例化部署和微服务间存在的依赖关系,使得准确捕捉微服务调用拓扑与微服务时延之间的关联性十分困难.针对该问题,提出了一种基于门控时空卷积网络(gated attention spatio-temporal graph convolution-al network,GaSTGCN)的时延预测模型.通过构建微服务拓扑调用实例的图数据序列,采用多层时空卷积图神经网络模型,在考虑微服务调用拓扑的同时,捕捉微服务节点时间和空间上特征的关联性;考虑微服务应用规模日渐庞大、时空关联性更难捕捉的现状,结合门控卷积模型,采用膨胀卷积技术与自适应门控机制,更精确地获取局部与全局微服务依赖特征.实验表明,所提模型具有较好的收敛性能,并且预测精度优于传统的预测算法.
Abstract
The multi-instance deployment of microservices and their interdependencies make it challenging to accurately capture the relationship between microservice invocation topology and service delay.To address this issue,a delay predic-tion model based on the gated attention spatial-temporal graph convolutional network(GaSTGCN)is proposed.By construc-ting graph data sequences of microservice invocation instances,a multi-layer spatial-temporal graph convolutional network is employed to capture temporal and spatial feature correlations of microservice nodes while considering the service invocation topology.Recognizing the increasing complexity of microservice application scales and the difficulty in capturing spatial-tem-poral correlations,the model integrates gated convolution mechanisms with dilated convolution techniques and adaptive ga-ting mechanisms to more accurately extract local and global microservice dependency features.Experimental results demon-strate that the proposed model achieves superior convergence performance and prediction accuracy compared to traditional prediction algorithms.
万方数据