Performance Management of Web Services Using Machine Learning Techniques

Abstract

Delivering fast response times for user transactions is a critical requirement for Web services. Long response times can frustrate users and can cause them to discontinue using the service. Therefore, performance plays an important role in the success of a Web service. This thesis focuses on two problems related to the performance of Web services. The first problem considers Web services that rely on cloud platforms. The virtualization technologies employed by clouds can trigger contention between virtual machines, thereby leading to long response times for the Web service. This necessitates the need for automated techniques that can detect and mitigate such contentions at runtime. The second problem focuses on developing a performance prediction technique that can provide insights on how to provision the right amount of resources to the Web service. Web services can experience a variety of workloads depending on the number and the usage pattern of their end users. Operators need a systematic technique for predicting the performance of a workload under a given resource allocation. An overarching aspect of both problems is the need to predict the performance of a Web service. Most of the existing prediction techniques require an expert to devise a performance model of the system. This thesis focuses on exploring a model-free data-driven approach. Historical performance logs of a Web service in combination with a machine learning approach is used. The use of the machine learning algorithm helps in providing predictions for workloads not directly observed in past. Results show that the proposed approach is more effective at contention detection than the baseline techniques and the output can be used to mitigate performance contention effectively. Results for response time percentile prediction show that the designed technique can predict response time percentiles with high accuracy for a wide variety of workloads. This technique can accurately estimate the resources needed to maintain response time percentiles within operator-specified thresholds.