User-Level Profiler-Based Predictive Framework for Applications in Multi-Tenant Clouds




Moradi, Hamidreza

Journal Title

Journal ISSN

Volume Title



Cloud computing has been adopted by many organizations as their main computing infrastructure due to its low cost of ownership and flexible resource management. However, applications running on the clouds usually share hardware resources with other virtual machines (VMs) and applications from other cloud users/tenants. Such hardware resource sharing among multiple tenants causes resource contention, which in turn degrades the performance of applications running on clouds. Moreover, the source of contention can vary due to changes of co-located VMs and their applications, which makes a target cloud application experience uncontrolled performance variations and fluctuations at runtime. However, to maximize the cost benefits of cloud deployments with optimal resource allocation, or to satisfy the timeliness requirements of time-sensitive applications, cloud users may need to have an accurate knowledge of the performance of their applications. This dissertation aims to provide novel frameworks and techniques for performance modeling and prediction of cloud applications from the perspective of ordinary cloud users. The proposed frameworks use several micro-benchmarks to measure the severity of resource contention in available system resources in a multi-tenant cloud environment. Then, based on the execution of an application, the effect of resource contention on the application's performance can be modeled for future predictions. First, we investigate the required micro-benchmarks and their settings thoroughly for accurate contention estimation of system resources in clouds. This includes identifying the minimum subset of system resources to profile and profiling duration that minimizes the overhead while providing accurate results. Second, to assess the feasibility of performance prediction from the cloud users' perspective, we designed uPredict, a User-level Profiler-based Predictive Framework in Multi-Tenant Clouds. uPredict can predict cloud applications' performance with high accuracy using sophisticated machine-learning models and the devised micro-benchmarks. Third, we extend the uPredict framework to an online User-level Adaptive Performance Modeling and Prediction framework (User-APMP). User-APMP will update the applications' performance model after cloud deployment with newly observed data to achieve higher accuracy at runtime. For updating the models both periodic model retraining and progressive model training have been considered. Fourth, we extend the framework to model and predict the performance of cloud applications executing on multiple VMs using distributed and hierarchical performance modeling and prediction framework. From the practical standpoint, our findings can provide valuable guidelines for modeling cloud-running applications, helping ordinary cloud users to reduce their cloud usage costs, and efficiently use acquired resources.


This item is available only to currently enrolled UTSA students, faculty or staff. To download, navigate to Log In in the top right-hand corner of this screen, then select Log in with my UTSA ID.


Cloud Computing, Machine Learning, Performance Modeling



Computer Science