Abstract

In this paper we introduce techniques that support advance resource reservation and admission control for applications acquiring information technology (IT) resources from resource utilities. These resource utilities offer programmatic access to resources for complex multi-tier applications. Such utilities may participate in grids. As a workload, we consider business applications which require resources continuously but that have resource demands that change regularly based on calendar patterns such as time of day and day of week. Applications acquire resources as needed to ensure a quality of service to their end users and they release resources when they are not needed to lower their infrastructure costs. We characterize the resource demands of such applications statistically using application demand profiles. The profiles are used to make resource reservations. An admission control system exploits the profiles to enable the overbooking of resources while offering statistical assurances regarding access to resources. Different assurance levels correspond to alternative classes of service. Policing techniques determine whether requests for resources conform to a reservation and therefore whether they must be serviced. We illustrate the feasibility of our approach with a case study that uses resource utilization information from 48 data center servers. Simulation experiments explore the sensitivity of the assurances to correlations between application resource demands, the precision of the demand profiles, and the effectiveness of the policing mechanisms.

Keywords: Resource management; Grid computing; Utility computing; Business applications

Article Outline

1. Introduction

2. Related work

3. Application demand profiles and assurances

3.1. Application demand profiles

3.2. Statistical assurance

3.2.1. Worst-case bound

3.2.2. Measurement approach

4. Resource access classes of service

5. Policing and entitlements

6. Case study

6.1. Measurement data

6.2. Application demand profiles

6.3. Utility ADP

6.4. Sensitivity analysis

6.4.1. Sensitivity of Γ to correlations between application demands

6.4.2. Validation by simulation

6.4.3. Sensitivity to the precision of ADPs (by simulation)

6.5. Policing

7. Summary and conclusions

Acknowledgements

References