Host Prediction And Auto-Scaling For Cloud-Based Load Testing

Modern applications are often deployed on the clouds,which enables users to obtain services via browsers or other clients.This adds convenience and openness to services but brings more uncertainty on their performance and usability.Load testing is an important way of evaluating application performance,which can effectively reduce performance risks and improve software quality.Traditional load testing is often conducted on local clusters,where workloads are created by the cooperation of multiple hosts to accomplish testing tasks.Local clusters need professional technicians to maintain.The hardware costs of local clusters are also high.It is hard to expand hardware resources when load testing needs more hosts to initiate load requests.Cloud-based load testing manages and performs test tasks on the clouds,and the workloads are initiated from virtual hosts.In this way,test resources can be easily expanded without manually deploying test clusters.It is convenient to perform cloud-based load testing,and the use of test resources could be more optimized.However,there yet lacks effective methods to predict the number of hosts needed by load testing.Non-enough host resources would cause the failure of load testing,while unnecessary more hosts would cause resources wastes.Besides,there are also resource schedule problems during the process of cloud-based load testing.The existing approaches are challenging to schedule load testing resources in time and the necessary amount.It is hard to ensure the quality of load testing.To solve the problems,we present a host prediction method for cloud-based load testing,which predicts the computing resources needed to complete the load testing according to the target testing scale and then converts the computing resources into a specific number of hosts.We also propose a technique to auto-scale the test resources according to certain load change strategies.The technique makes host allocation plans to allocate appropriate numbers of hosts before the designated load scales to ensure that workload can always be initiated with enough computing resources.It can reduce the host allocation delay,minimize resources wastes,and optimize the use of resources during the testing process.In more detail,the work in this thesis includes the follows.(1)We propose a host number prediction method for cloud-based load testing.It firstly conducts small-scale load testing to collect data about the workloads and the used computing resources.We then use machine learning to train a model to predict the computing resources required to initiate certain scales of workloads.The method estimates the number of demanded hosts from the predicted amount of computing resources.With these hosts,load testing can be effectively conducted,and resources wastes can be largely avoided.(2)We present a technique to auto-scale the test resources according to a load change strategy.The technique uses historical host initialization time to train a model to predict the preparation time needed for creating certain number of new hosts.We predict the number of hosts required for each stage of load testing according to the load change strategy.According to the number of the required hosts and the host preparation time,we estimate the time to start and shutdown virtual hosts so that the test resources can be auto-scaled,and the load testing can be effectively conducted.(3)Finally,we implement the host prediction and auto-scaling technique introduced in this thesis and integrate them into a cloud testing tools.We conduct experiments to demonstrate the effectiveness of the proposed methods.These methods can help optimize the use of computing resources and reduce test costs.