Provisioning at a Utility Computing Site: The Resource Assignment Problem

In our view, Utility Computing is the computing model of the future. There will exist different kinds of computing utilities that provide computing resources and capabilities as a service. Typically these computing utilities will have a complex infrastructure, which needs to be scalable, economical, self-adapting, trusted, and federated. HP's Utility Data Center (UDC) is an example, which aims at providing a power plant that enables utility computing. A UDC can contain tens of thousands of servers and storage devices connected through a shared high-speed network fabric. Managing both the infrastructure and the applications in such a large environment raises many challenges that do not exist in today's data centers. To increase asset utilization and achieve economies of scale, most physical resources will be virtualized and shared across multiple applications. Application services need to be provisioned automatically and deployed in a timely fashion. More specifically, the main focus of UDCs is to provide infrastructure support for business applications, which typically have continuous but time-varying demand on resources. Once an application is admitted into the UDC, the resources it requires should be assigned programmatically and efficiently.

Click to view larger version of this image.		This project is centered around models and solution algorithms for the automated assignment of computing resources to applications. We take into account technological constraints imposed by the topologies of applications, as well as the infrastructure, while optimizing the assignments based on quality of service and cost objectives.

The problems are challenging due to their size and complexity. Our solution approaches range from quadratic and linear Mixed Integer Programming formulations to efficient heuristics.

Our preliminary results show that we can solve the resource assignment problems with thousands of servers and hundreds of applications in a couple of minutes. This technology is part of the HPL adaptive infrastructure initiative and will be a component of the resource provisioning system prototype that HPL is building.

For more detailed technical information on one of the approaches please refer to Technical Report HPL-2002-64R1.

Contacts:
Project Manager: Dirk Beyer (650)-236-2711
Project Lead: Cipriano (Pano) Santos (650)-857-2476