Performance Management of IT Infrastructure

_Principal Investigator:_
Johnny Wong, Professor
David R Cheriton School of Computer Science, University of Waterloo

Background: The proposed research is concerned with autonomic resource management for an IT infrastructure which consists of a variety of computing resources, e.g., clusters, individual servers, and grid. These resources may be geographically distributed and are accessed via a virtualization layer. A diverse set of applications will be supported. These applications have varying workloads and performance requirements. There is also a cost associated with resource usage. Resource management for such a computing environment is a challenging task. In our proposed research, resource allocation algorithms that provide autonomic capabilities for configuration, reconfiguration and tuning will be investigated.

Objectives: In our investigation, virtualization is used as an enabling technology for resource management. Virtualization provides a unified view of an organization’s computing resources. These resources are given the appearance of a centralized data centre.

We first consider a scenario where a pool of processor nodes is shared by two or more clusters. These clusters execute jobs belonging to different classes, e.g., interactive jobs and batch jobs. For each cluster, the cost of using a processor node is modeled as a time-varying cost function. The workload of each job class may fluctuate over time. We will develop and evaluate algorithms which dynamically allocate processor nodes to each cluster such that usage cost is kept to a minimum and the service level agreements for the various job classes are met. We then extend our investigation to an IT infrastructure that consists of clusters, individual servers, and grid. There may be constraints on the amount of resources to be added or removed each time a resource reallocation decision is made. Policy on resource usage and the ability to do advance resource reservation will also be included in our investigation.

Our algorithms operate as follows. Resource allocation decisions are made at decision points. At each decision point, a predictive performance model is used to determine the amount of resources that should be allocated to each job class, using as input the job profiles, cost functions, performance requirements, and measurement data on workload and performance parameters. This model is solved by a combination of analysis and simulation. Resource reallocation will then be guided by the amount of resources that should be allocated to each job class, as determined by the model. We plan to evaluate the effectiveness of our algorithms using an experimental facility to be established at the University of Waterloo, and by simulation. We also plan to explore the use of the IBM Tivoli Intelligent Orchestrator for performance monitoring and autonomic resource provisioning.

Potential benefit to Ontario: Our algorithms will contribute to the design of a virtual infrastructure that exhibits a high degree of autonomic capabilities. Once deployed, the applications will function within the parameters defined by service level agreements. Resource allocation will be adjusted by autonomic managers by self-tuning, self-balancing, and cost considerations. This can potentially lead to a reduction in the time and effort in managing IT infrastructures. Such cost savings would be a potential benefit to firms and organizations in Ontario. This project will also train highly qualified personnel who are well prepared to provide technical leadership in autonomic resource management and capacity planning. In addition, our results are potentially useful when a company wishes to design tools for autonomic resource management.

_Other Projects_