Adapting Server-Resource Assignments in Data Centers Based on Time-Interval Clustering (Abstract Only)

Kurzfassung/Abstract

In today's virtualized data centers, energy consumption of running servers is one of the key cost drivers. Typically, enterprise applications are hosted in virtual machines (VM), which are co-hosted on the same physical server to reduce the server count. Live-migration of VMs can lead to additional savings by temporarily shrinking the server park. For example, at night time workloads of most applications are lower and a smaller number of servers would be sufficient to host the applications. On the other hand, migrations require a significant amount of CPU and memory resources on the physical servers and administrators prefer stable allocations, which require reassignments only in a prespecified number of time-periods. We cluster features from large volumes of workload traces, which allows for determining (near-) optimal time periods where reassignments are beneficial. The approach is evaluated in numerical experiments based on workload data from a professional data center.