Motivation
Enterprises host a multitude of internal business applications,
e.g., for customer relationship management (CRM), sales and
distribution, and accounting in their data centers. The
underlying system consisting of hardware and software if often
sized according to expected peak loads in order to meet service
level objectives like availability and performance. However, as
the load on the data center is typically volatile, the systems are
over-provisioned, causing high total costs of ownership, e.g.,
high investment, maintenance, and energy costs. Furthermore,
changing demands and hardware failures require continuous
adaptations of the IT infrastructure and result in high
administration costs and operational hazard.
By dynamically allocating hardware resources, an enterprise can
significantly reduce its operating costs and administration
expenses. In this context, virtualization constitutes an
important technique. It allows for flexible and efficient resource
allocation in data centers. Today, humans control the virtual
machines when, e.g., hardware or software failures occur, load
peaks exceed the system capacity, or additional resources are
provided for a virtual machine. The administrator must find the
correct action to take based on monitored data. We are aiming at
automating the management of such data centers by, e.g.,
automatically controlling resource allocations and virtual machine
migrations.
Research topics
- Automated, adaptive control mechanisms for virtualized data centers
- Enhancement of data center operations (minimizing operational costs for running and cooling servers) while satisfying resource and applications level service level agreements
- Development and empirical evaluation/comparison of different control mechanisms for efficient and adaptive resource allocation for enterprise applications
- Adaptivity refers to
- Providing a hosting environment that supports large sets of highly scalable applications
- The ability to deal with hardware and software failures
- Methods from time series analysis (forecasting, classification and pattern recognition), operations research (bin packing, scheduling) and information systems
- Main challenges include the identification of appropriate control mechanisms for different types of workloads