%0 Thesis %1 Herbst2012a %A Herbst, Nikolas Roman %C Am Fasanengarten 5, 76131 Karlsruhe, Germany %D 2012 %K Award Cloud Online_monitoring_and_forecasting Prediction Resource_management Statistical_estimation_and_machine_learning Thesis_supervised_by_Nikolaus_Huber Thesis_supervised_by_SE_member Tool WCF descartes se2_mastersthesis t_studentthesis %T Workload Classification and Forecasting %X Virtualization technologies enable dynamic allocation of computing resources to execution environments at run-time. To exploit optimisation potential that comes with these degrees of freedom, forecasts of the arriving work's intensity are valuable information, to continuously ensure a defined quality of service (QoS) definition and at the same time to improve the efficiency of the resource utilisation. Time series analysis offers a broad spectrum of methods for calculation of forecasts based on periodically monitored values. Related work in the field of proactive resource provisioning mostly concentrate on single methods of the time series analysis and their individual optimisation potential. This way, usable forecast results are achieved only in certain situations. In this thesis, established methods of the time series analysis are surveyed and grouped concerning their strengths and weaknesses. A dynamic approach is presented that selects based on a decision tree and direct feedback cycles, capturing the forecast accuracy, the suitable method for a given situation. The user needs to provide only his general forecast objectives. An implementation of the introduced theoretical approach is presented that continuously provides forecasts of the arriving work's intensity in configurable intervals and with controllable computational overhead during run-time. Based on real-world intensity traces, a number of different experiments and a case study is conducted. The results show, that by use of the implementation the relative error of the forecast points in relation to the arriving observations is reduced by 63% in average compared to the results of a statically selected, sophisticated method. In a case study, between 52% and 70% of the violations of a given service level agreement are prevented by applying proactive resource provisioning based on the forecast results of the introduced implementation.

@mastersthesis{Herbst2012a, abstract = {{Virtualization technologies enable dynamic allocation of computing resources to execution environments at run-time. To exploit optimisation potential that comes with these degrees of freedom, forecasts of the arriving work's intensity are valuable information, to continuously ensure a defined quality of service (QoS) definition and at the same time to improve the efficiency of the resource utilisation. Time series analysis offers a broad spectrum of methods for calculation of forecasts based on periodically monitored values. Related work in the field of proactive resource provisioning mostly concentrate on single methods of the time series analysis and their individual optimisation potential. This way, usable forecast results are achieved only in certain situations. In this thesis, established methods of the time series analysis are surveyed and grouped concerning their strengths and weaknesses. A dynamic approach is presented that selects based on a decision tree and direct feedback cycles, capturing the forecast accuracy, the suitable method for a given situation. The user needs to provide only his general forecast objectives. An implementation of the introduced theoretical approach is presented that continuously provides forecasts of the arriving work's intensity in configurable intervals and with controllable computational overhead during run-time. Based on real-world intensity traces, a number of different experiments and a case study is conducted. The results show, that by use of the implementation the relative error of the forecast points in relation to the arriving observations is reduced by 63% in average compared to the results of a statically selected, sophisticated method. In a case study, between 52% and 70% of the violations of a given service level agreement are prevented by applying proactive resource provisioning based on the forecast results of the introduced implementation.}}, added-at = {2020-04-08T16:48:07.000+0200}, address = {Am Fasanengarten 5, 76131 Karlsruhe, Germany}, author = {Herbst, Nikolas Roman}, biburl = {https://www.bibsonomy.org/bibtex/261c15717cb0160b93e2ea1cec6eb81ce/se-group}, interhash = {3263a81db3cb330219208a0763fa543e}, intrahash = {61c15717cb0160b93e2ea1cec6eb81ce}, keywords = {Award Cloud Online_monitoring_and_forecasting Prediction Resource_management Statistical_estimation_and_machine_learning Thesis_supervised_by_Nikolaus_Huber Thesis_supervised_by_SE_member Tool WCF descartes se2_mastersthesis t_studentthesis}, note = {<b>Forschungszentrum Informatik (FZI) Prize "Best Diploma Thesis"</b>}, school = {{Karlsruhe Institute of Technology (KIT)}}, timestamp = {2020-10-20T11:34:02.000+0200}, title = {{Workload Classification and Forecasting}}, type = {{Diploma Thesis}}, year = 2012 }