Workload modeling and prediction for resources provisioning in cloud
The evaluation of resource management policies in cloud environments is challenging since clouds are subject to varying demand coming from users with different profiles and Quality de Service (QoS) requirements. Factors as the virtualization layer overhead, insufficient trace logs available for anal...
| Autor: | |
|---|---|
| Tipo de documento: | tese |
| Estado: | Versão publicada |
| Data de publicação: | 2017 |
| País: | Brasil |
| Recursos: | Universidade Federal do Ceará (UFC) |
| Repositório: | Repositório Institucional da Universidade Federal do Ceará (UFC) |
| Idioma: | inglês |
| OAI Identifier: | oai:repositorio.ufc.br:riufc/22987 |
| Acesso em linha: | http://www.repositorio.ufc.br/handle/riufc/22987 |
| Access Level: | Acceso aberto |
| Palavra-chave: | Teleinformática Modelagem computacional Simulação (Computadores) Computação em nuvem Cloud computing Simulation Workload modeling Workload profiling |
| Resumo: | The evaluation of resource management policies in cloud environments is challenging since clouds are subject to varying demand coming from users with different profiles and Quality de Service (QoS) requirements. Factors as the virtualization layer overhead, insufficient trace logs available for analysis, and mixed workloads composed of a wide variety of applications in a heterogeneous environment frustrate the modeling and characterization of applications hosted in the cloud. In this context, workload modeling and characterization is a fundamental step on systematizing the analysis and simulation of the performance of computational resources management policies and a particularly useful strategy for the physical implementation of the clouds. In this doctoral thesis, we propose a methodology for workload modeling and characterization to create resource utilization profiles in Cloud. The workload behavior patterns are identified and modeled in the form of statistical distributions which are used by a predictive controller to establish the complex relationship between resource utilization and response time metric. To this end, the controller makes adjustments in the resource utilization to maintain the response time experienced by the user within an acceptable threshold. Hence, our proposal directly supports QoS-aware resource provisioning policies. The proposed methodology was validated through two different applications with distinct characteristics: a scientific application to pulmonary diseases diagnosis, and a web application that emulates an auction site. The performance models were compared with monitoring data through graphical and analytical methods to evaluate their accuracy, and all the models presented a percentage error of less than 10 %. The predictive controller was able to dynamically maintain the response time close to the expected trajectory without Service Level Agreement (SLA) violation with an Mean Absolute Percentage Error (MAPE) = 4.36%. |
|---|