An Effective Channel and Spatial Attention based Parallel Convolutional Model for Micro-Service System Anomaly Multi-Classification

Micro-service architecture is a popular architectural style and approach in developing large-scale applications in the cloud due to its high availability, low coupling, and easy scalability. To ensure its Quality of Service (QoS), it is crucial to efficiently diagnose the runtime system anomalies. However, typical approaches are one-class anomaly detection, which is inadequate to identify specific system anomaly types of micro-services in the cloud environment. In addition, it is significantly challenging to effectively classify the runtime anomalies due to the complex dependencies among various micro-services and the high dynamism of the cloud environment. In this paper, we propose a Channel and Spatial Attention based Parallel Convolutional System Anomaly Multi-Classification (CSA-PC-SAMC) in micro-service architecture. To achieve an accurate and robust anomaly classification, we construct a parallel convolutional architecture that allows subnetworks to extract features independently. Additionally, we incorporate channel and spatial attention in the parallel convolutional layers to mitigate the loss of feature representation. Empirical experiments have been conducted on the micro-service benchmark Sock-Shop with three types of system anomalies: CPU hog, memory leaks, and network delay. The experimental results revealed that the average Macro F1 and Micro F1 scores were 0.909 and 0.972, respectively, indicating that CSA-PC-SAMC achieved the top ranking in the comprehensive evaluation, compared with other models, it has increased by 37.9% and 4.4% respectively.