Forecasting methods for cloud hosted resources, a comparison

Van Greunen, Manrich (2015-12)

Thesis (MSc)--Stellenbosch University, 2015.

Thesis

ENGLISH ABSTRACT: Cloud computing has revolutionised the modern day IT industry and continues to foster the development of new products and services. Amid the dynamically changing workloads presented to cloud computing lies the challenge of ensuring sufficient resources are available when needed. Recently, proactive provisioning and auto-scaling schemes have emerged as solutions to this. Forecasting methods are inherent to these provisioning schemes and to the author's knowledge, no formal investigation has been performed in comparing different forecasting methods. The purpose of this research was to investigate various forecasting methods presented in recent research, adapt evaluation metrics from literature and compare these methods on prediction performance using two real-life cloud resource datasets. It was found that less complex methods, such as moving average and autoregression outperformed other more complex methods that were investigated, on the majority of used evaluation metrics. We also found that our 30th order auto-regression model achieved statistically significantly better results compared to the other forecasting methods. Furthermore, there was no single evaluation metric that gave concise comparative results between forecasting methods, but overload likelihood ratio as metric showed great promise to this end. It was argued that focus should be put on developing evaluation metrics that specifically relate to the cloud environment and further investigation should be performed on a closed-loop system or real-life cloud platform. Cloud computing has become ubiquitous with the Internet as we know it today. We believe that effective provisioning of cloud computing resources should be at the core of modern cloud management systems and the primary objective of cloud platform providers.

AFRIKAANSE OPSOMMING: Die wolk-verwerking revolusie in hedendaagse IT industrieë ontwikkel voortdurend nuwe produkte en dienste. Te midde van die dinamiese gedrag van wolk verwerking, as gevolg van veranderende werkslandings op wolke, is dit 'n uitdaging om te verseker dat genoeg verwerker-hulpbronne beskikbaar is voordat dit benodig word. Ontwikkelinge in pro-aktiewe voorsiening en outomatiese skallerings skemas was onlangs gemaak ter oplossing vir hierdie uitdaging. Inherent aan hierdie skemas is die gebruik van vooruitskattingsmetodes en sover die outeur se kennis strek, is daar tans geen resultate van formele ondersoeke in die vergelyking van verskeie vooruitskattingsmetodes, beskikbaar nie. Die doel van hierdie navorsing was om ondersoek in te stel aangaande verskeie vooruitskattingsmetodes en die aanpas van evalueringsmaatstawwe soos genoem in literatuur. Met behulp van werklike wolk hulpbron datastelle was hierdie metodes met mekaar vergelyk. Daar is gevind dat eenvoudige metodes, soos gly-gemiddeld en outo-regressie, uitgeblink het wanneer dit gemeet was met die meerderheid van die maatstawwe. Ons 30ste orde outo-regressie model verkry die hoogste akkuraatheid. Verder, is daar gevind dat geen een evaluasie maatstaf 'n duidelike verskil tussen metodes uitwys nie, maar dat die oorbelas waarskynlikheidsverhouding vir hierdie doel belowend lyk. Daar is aangevoer dat fokus geplaas moet word op die ontwikkeling van evalueringsmaatstawwe wat spesifiek verwant is aan die wolk omgewing en verdere ondersoek op 'n geslote-lus stelsel of werklike wolk platform, gedoen moet word. Wolk-verwerking is alomteenwoordig met die Internet soos ons dit vandag ken. Effektiewe voorsiening van wolk hulpbronne en die gebruik van vooruitskattingsmetodes is die kern van moderne wolk bestuurstelsels. Wolk platform verskaffers behoort dit as hul primêre doel tot sukses te beskou.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/98032
This item appears in the following collections: