VM and workload fingerprinting for software defined datacenters

Abstract

In this work we develop strategies for mining telemetry streams in virtualized clusters to automatically discover relationships between sets of virtual machines. Examples of relationships include correlations between virtual machines, similarities in resource consumption patterns or dominant resources, and similarities in metric variations. The main challenge in our approach is to transform the raw captured data consisting of resource usage and VM-related metrics into a meaningful fingerprint that identifies the virtual machine and describes its performance. In our analysis we try to determine which of these metrics are relevant and how they can be expressed as a light-weight and robust fingerprint that offers insight about the status of the machine.