The Cost Advantage of Virtual Machine Migrations: Empirical Insights into Amazon's EC2 Marketspace

TL;DR

This paper analyzes the cost benefits of creating heterogeneous virtual machine portfolios and migrating VMs across marketplaces in Amazon EC2, highlighting significant potential for cost optimization based on empirical data.

Contribution

It provides an empirical cost analysis of cloud portfolios and migration strategies, revealing optimal practices for cost savings in cloud resource management.

Findings

Heterogeneous portfolios reduce costs compared to homogeneous ones.

VM migration during runtime can be cost-effective for certain durations.

Most VM resources are underutilized, indicating high potential for cost optimization.

Abstract

In recent years, cloud providers have introduced novel approaches for trading virtual machines. For example, Virtustream introduced so-called muVMs to charge cloud computing resources while other providers such as Google, Microsoft, or Amazon re-invented their marketspaces. Today, the market leader Amazon runs six marketspaces for trading virtual machines. Consumers can purchase bundles of virtual machines, which are called cloud-portfolios, from multiple marketspaces and providers. An industry-relevant field of research is to identify best practices and guidelines on how such optimal portfolios are created. In the paper at hand, a cost analysis of cloud portfolios is presented. Therefore, pricing data from Amazon was used as well as a real virtual machine utilization dataset from the Bitbrains datacenter. The results show that a cost optimum can only be reached if heterogeneous portfolios are created where virtual machines are purchased from different marketspaces. Additionally, the cost-benefit of migrating virtual machines to different marketplaces during runtime is presented. Such migrations are especially cost-effective for virtual machines of cloud-portfolios which run between 6 hours and 1 year. The paper further shows that most of the resources of virtual machines are never utilized by consumers, which represents a significant future potential for cost optimization. For the validation of the results, a second dataset of the Bitbrains datacenter was used, which contains utility data of virtual machines from a different domain of application.