Virtual Desktop User Logon Behavior Analysis Via Machine Learning

Virtual Desktop Infrastructure(VDI)is a virtualization technology that hosts the desktop operating system on centralized servers in a private or public cloud data center and isolates the desktop environment from the physical client devices used to access the desktop.Effective resource management is very important for virtual desktop systems.On the one hand,the system needs to power-on enough virtual desktops to ensure user experience,and on the other hand,it needs to power-off as many idle virtual desktops as possible to save running costs.The current VDI systems manage resources in a passive way by either reactive capacity tuning or manual configuration according to user needs.This dissertation focuses on innovative in-depth research on virtual desktop intelligence to analyze virtual desktop user logon behavior data.Based on the different resource management characteristics of different types of desktop pools,specific advanced machine learning models are induced to predict the future system work load accurately.As the result,the resource management efficiency of the virtual desktop system can be greatly improved,and the operation cost can be effectively reduced on the premise of ensuring the user experience.The major contributions of this dissertation can be summarized as follows:1.The non-persistent desktop pools do not assign specific desktops to users.Based on this characteristic,a pool-level workload prediction model CAFE(Co Arse to Fine historical d Escriptive)is proposed.Specifically,CAFE defines multiple extraction levels according to the time distance and transforms the aggregate session count with different granularities at different levels.Accordingly,the extracted multi-grained historical features can describe the overall user logon behavior from different aspects.These features serve as the training data of the ensemble learner to obtain an accurate workload prediction model.Experimental results on real customer datasets demonstrate the effectiveness of the CAFE model for workload prediction of the nonpersistent desktop pools.2.The persistent desktop pools allocate a dedicated desktop for each user.Therefore,resource management of the persistent pools must be precise down to the single desktop level which requires the the accurate prediction of the single-user logon behavior.Accordingly,a single-user logon prediction model SOUP(Single-user log On behavio Ur Prediction using encoded multi-grained features)based on encoded multi-grained features is proposed.The SOUP model describes the historical logon behavior of each user as time series data with values of0 and 1,and encodes the seasonal features via the coarse-to-fine multi-granularity representation.Furthermore,SOUP model combines all user data in a persistent pool to train a general pool-level model,which can mitigate the impact of data noise and lazy users.The experimental results on real customer datasets show that the SOUP model can achieve superior performance on the single-user logon behavior prediction with much lower model training cost.3.The above single-user logon behavior prediction model is based on SISL(SingleInstance Single-Label learning)framework where the close relationships among the VDI poolsharing users are not considered.Accordingly,an innovative approach BAMBOO(BAlanced Miml BOOst)is proposed to model VDI user logon behavior using the MIML(Multi-Instance Multi-Label learning)framework.Specifically,BAMBOO groups each user with selected supporting users to jointly model their logon behaviors with multi-instance representation in the feature space and multi-label prediction in the output space.The MIML model is then optimized by adding balanced error-rate minimization in the MIMLBoost procedure.The experimental results demonstrate that the BAMBOO model outperforms the state-of-the-art VDI single user logon behavior prediction models.4.From engineering practice perspective,the overall architecture of the machine learning based VDI proactive resource management system using CAFE,SOUP and BAMBOO models is proposed.Specifically,the overall architecture is composed of the desktop management subsystem and the workload prediction service based on Amazon AWS related services and technologies.The desktop management subsystem adds the data sending module and proactive resource management service to the existing resource management system.The workload prediction subsystem has a three-layer structure,including the service layer,machine learning pipelines and the data layer.The experimental deployment of the prototype system shows that the machine learning based VDI proactive resource management system can significantly reduce the operating cost of the VDI desktop pool without impacting the user experience.Six chapters are included in this thesis.The first chapter introduces the background and research status of virtual desktop user behavior analysis,and summarizes the work of this thesis.The second chapter introduces the technical details of the CAFE model and the comparative experiments based on real-world VDI data and open power data.The third chapter introduces the technical details of the SOUP model and the comparative experiments based on the real-world persistent desktop pool dataset,and also summarizes the multi-granularity feature extraction method.Chapter 4 introduces the technical details of the BAMBOO model and related comparative experiments.Chapter 5 introduces the overall architecture and the application results of the machine learning based VDI proactive resource management system.Finally,Chapter 6summarizes the main contributions of this paper and gives an outlook on future work.