Scholarlink

Applications are divided into a set of small, loosely coupled microservices, each of which contains multiple container replicas deployed on different Edge nodes. As users move around, they send requests to microservices, and causing workload fluctuations. Therefore, there is a need of workload prediction for microservice containers. However, most of the existing workload prediction methods only take into account the relationship among the Edge nodes and ignore the dependencies among these microservices. In this paper, a Multi-graph Embedding and GRU Based workload Prediction (MGEBP) method for microservices is proposed. Firstly, pairwise correlations of geographical dependencies of Edge nodes and dependencies among microservices are encoded into multiple graphs, which are embedded into a vector. Secondly, recurrent neural networks are utilized to learn temporal dependencies. Experiments show that the Root Mean Square Error (RMSE) of the MGEBP model is about 5% lower than that of the state-of-art algorithm.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-graph Embedding and GRU Based Prediction for Mesh Microservice in Edge Computing

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Mengyu Chen,
  • Zhicheng Cai

摘要

Applications are divided into a set of small, loosely coupled microservices, each of which contains multiple container replicas deployed on different Edge nodes. As users move around, they send requests to microservices, and causing workload fluctuations. Therefore, there is a need of workload prediction for microservice containers. However, most of the existing workload prediction methods only take into account the relationship among the Edge nodes and ignore the dependencies among these microservices. In this paper, a Multi-graph Embedding and GRU Based workload Prediction (MGEBP) method for microservices is proposed. Firstly, pairwise correlations of geographical dependencies of Edge nodes and dependencies among microservices are encoded into multiple graphs, which are embedded into a vector. Secondly, recurrent neural networks are utilized to learn temporal dependencies. Experiments show that the Root Mean Square Error (RMSE) of the MGEBP model is about 5% lower than that of the state-of-art algorithm.