DiagSoftfailure: Automated Soft-Failure Diagnostic Tool Using Machine Learning for Network Users

As increasing individuals and organizations move their activities and services into online, network performance problems resulting in slow data communication speed becomes the significant obstacle for satisfactory user experience. Currently, there is a lack of a fully automated tool that can help network users to find the complicated network problems that degrade the performance of network applications. To fill this gap, we proposed an Automated Soft- Failure Diagnostic Tool Using Machine Learning (DiagSoftfailure) for Network Users to infer the location and root cause of network failures that result in performance degradation. Upon the user’s request via the web browser, DiagSoftfailure sever (DS) deployed at the border router between the campus/enterprise LAN and the backbone network collects and analyzes the packet trace corresponding to the target application for soft-failure diagnosis. It first utilizes open source TCP trace analysis software (i.e, libpcap and tcptrace) to obtain the raw features of the network behavior. Then, those raw features are further processed to extract network signatures that can provide sufficient distinction for an effective and reliable diagnosis. Base on the network signature, automated classifiers trained by combining supervised and semi-supervised machine learning is used to identify both known and unknown soft-failures in the network. Finally, the DS sends the network user a diagnosis report which allows novice and expert users to view and understand network condition and problems. DiagSoftfailure provides the capabilities for automated network soft-failure diagnosis with the following features: (i) a user-focused diagnosis requires no cooperation with the network manager; (ii) an adaptive network signature that is robust against data inconsistency and high-dimensionality of network behavior data ensures high diagnosis accuracy; (iii) capable of identifying unknown faults by combining supervised and unsupervised machine learning; (iv) requires no changes in OS system kernel and allows implementation flexibility. (v) a diagnosis report groups test results into the different categories in a comprehensive format and can be understood by novice users. DiagSoftfailure is designed as an automated user-focused network diagnosis tool that helps network users to actively find the performance problems that cause the application to run slower than expected. Once developed as COTS/GOTS products, it can significantly improve the user’s experience on network performance by helping them identifying and solving the network soft-failures: (1) Improve User Satisfaction: The DiagSoftfailure is designed to quickly and easily identify a specific set of conditions that impact the network performance. It can help users to find the root cause of the network performance degradation. It can assist the user and network administrator to rapidly resolve the network problem and improve connection speeds and alleviate user dissatisfaction; (2) Implementation Flexibility: Users can diagnose their network problems without cooperation with the network operator, which a capability unavailable in many network diagnostic tools. Thus it can be implemented as an application in the LAN of the users. This provides flexibility for the network operator to deploy the DiagSoftfailure system at any desired network location; and (3) Reducing Network Maintenance Cost: DiagSoftfailure provides an automatic data collection and analysis tool for users to automatically find the root cause of the performance problem in the network. By leveraging the automated diagnosis, it reduces the manpower required for network diagnosis as much as possible. Therefore it can largely reduce the cost of network operation and system maintenance.