Characterization of load tests in web applications through polynomial regression

Introduction: One of the challenges of the infrastructure that supports web applications is to guarantee the quality of service (QoS), so it is necessary to characterize the performance of this infrastructure for the planning and configuration of both system capacity and resource management. This is how one of the most outstanding methods to evaluate the performance of the web application infrastructure are load or stress tests, which seek to determine the behavior of a website or web service in terms of the number of sequential or simultaneous requests that it receives. emulate real client connections. Objective: In this article, we propose a tool that contributes to the characterization of load or stress tests of web applications through the use of machine learning techniques and specifically through the use of supervised learning models associated with the polynomial regression method. Method: Using Pratt's iterative pattern, a tool based on free software libraries and Python language technologies was developed, which supports both the execution of load tests and the determination of the degree of the polynomial (2 or 3), allowing the curve to be characterized with the time it takes for the server to respond or service a certain percentage of requests, after detecting outliers in the captured data. Results: Using the proposed tool, a case study was carried out on the website of the University of Cartagena to verify the functionality and usefulness of the proposed approach in which a total of 50 sequential requests were executed, from which the tool determined a polynomial model of degree 3 that obtained as a result for the training set an RMSE value of 0.992 and an R2 value of 0.994, while for the test set an RMSE value of 10.447 and an R2 value of 0.744 were obtained. Conclusions: The algorithm represents a contribution regarding to the reviewed state of the art works and respect to different tools for the execution of load tests, which in general terms do not use machine learning methods to model the performance of the infrastructure that supports the applications. web, limiting itself to the execution of sequential and concurrent tests. Specifically, the approach proposed in this article allowed to determine a polynomial that characterizes the mathematical relationship between the percentage of requests served and the response time obtained by the web server.