Understanding and Predicting VM Costs in the Multi-Cloud Landscape

The current trend for ICT infrastructure is largely based on new architectures and paradigms, such as Cloud Computing and Virtual Machines (VMs). However, the complicated and obfuscated nature of the price structure across different Cloud Service Providers (CSPs) creates significant challenges for organisations in the quest for improved cost effectiveness and vendor lock-in avoidance. Existing comparison tools currently possess set price data but lack dynamic forecasting capabilities for custom VM builds. Organizations also need a systematic way of forecasting virtual machine expenses in various cloud services based on certain technical specifications. The disjointed nature of the cloud's pricing complicates infrastructure decisions, making informed choices difficult. This lack of clarity creates vendor lock-in situations in such a way that the costs of migration incurred end up being excessively costly, following the organizations' having embedded themselves in the provider's environment. This thesis establishes a common system for generating hourly VM costs in prominent CSPs. The process entails retrieving the price information of 20 providers through the API-based and web scraping methods, normalization of inhomogeneous technical details, and configuring forecasting models for approximating costs for any random VM setting. The thesis provides a wide-ranging dataset consisting of virtual machines from European, American, and Asian providers and contains detailed specifications such as CPU architecture, memory and storage, GPU capabilities, and locations. The paper describes the methodology used for acquiring the data, the complexities of standardization faced, an exploratory data analysis of discovered price trends, and the development of machine learning models for price prediction. The study finds important price differences by provider and geography, wherein European providers have typically provided better CPU-intensive configurations. Memory capacity is the most powerful predictor of the price of VMs, followed by vCPU and storage capacity. Geographic concentration finds European providers having more similar price strategies than hyperscale providers such as AWS and Azure, which have more varied regional pricing. The work outlines how machine learning can efficiently abstract complex cloud pricing schemes and provide organizations with the necessary cost estimation tools across various suppliers, enabling informed infrastructure decisions. The patterns extracted at the provider-based and geographic level reveal how mindful provider selection by workload definition and geographic coverage can result in effective cost reductions. The developed analysis tool facilitates the comparison of different provider offerings and addresses a fundamental challenge in multi-cloud cost optimization. Future research should increase the provider coverage, include temporal pricing dynamics by utilizing time-series analysis, and build real-time price forecasting capabilities. Integration with cloud orchestration platforms can facilitate auto-cost-aware resource provisioning, and working with CSPs can enhance the data quality and completeness of the features.