基于多通道注意力图卷积网络的微服务分解

Multi-channel Attentional Graph Convolutional Networks for Microservice Extraction

为了解决功能、规模和复杂性不断增长的软件系统可能面临的可维护性和可扩展性等一系列软件开发和运维问题,微服务分解成为了目前研究的热点。现有的微服务分解主要是通过微服务的聚类,将单体系统划分为潜在的微服务候选。在微服务的自动化聚类中,基于图卷积网络(Graph Convolutional Network,GCN)的深度学习方法在特征学习方面取得了较好的效果,但是现有模型中缺乏对多通道信息的处理。针对该问题,提出一种基于多通道注意力图卷积网络的微服务分解方法MAGEMP。该方法使用多通道图注意力网络来学习不同强度的属性图和结构图节点之间的特征嵌入表示,再通过注意力机制获取不同通道嵌入表示的融合信息,最后综合聚类信息的联合学习框架获得高质量的微服务分解。在四个公开数据集上多角度验证该模型的有效性。与同类方法相比,MAGEMP方法提高了嵌入特征学习能力,在单体程序公开数据集上测试的功能性、模块性等性能方面取得了显著提升。

In order to solve a series of software development and operation and maintenance problems,such as maintainability and scalability,which may be faced by software systems with increasing functions,scale and complexity,microservice extraction is a hot problem currently.The existing work of microservice extraction is mainly to divide the monolithic program into potential microservice candidates through the clustering of microservices.The graph convolutional network(GCN)for automatic microservice extraction has been obtained potential results,but being lack of taking full advantage of multi-channel information.To solve the above problems,a monolith decomposition method using deep learning clustering based on multi-channel attention map neural network(MAGEMP)is proposed.Multiple independent graph convolutional networks are used to learn different interactions between topological graph and attribute graph nodes of different modal,and then the fusion of different embedding representations is further obtained through the attention mechanism,finally the joint learning framework integrating clustering information obtains high-quality microservice division.The validity of the model is verified from multiple angles on four public data sets.Compared with similar methods,MAGEMP improves the learning ability of embedded features,and significantly improves the performance of testing on the open data set of monomer programs,such as functionality and modularity.