|
Dara Kusic
Advisor: Nagarajan Kandasamy, Ph.D.
Abstract:
Distributed computing systems comprising numerous and networked hardware and software components host applications including online banking and retail commerce with stringent quality-of-service (QoS) requirements. To operate such systems effectively, multiple performance-related parameters to provision computing resources to applications and control power consumption must be dynamically tuned in order to meet service level agreements under dynamic operating conditions. As applications and systems increase in scale and complexity, meeting QoS requirements via manual tuning is not just tedious and error-prone, but also infeasible. To address these challenges, it is highly desirable that computing systems become largely autonomic, i.e., capable of managing themselves, given high-level objectives from administrators.
We propose to develop a fault-adaptive control framework that will continually tune parameters in a distributed computing system to anticipate changes in the external operating environment (e.g., time-varying workload) and respond to changes within the system (e.g., component failures). The developed framework will solve the problem of on-demand computing---an emerging resource provisioning model to efficiently host online services, where computing resources made available to online services only as needed, rather than statically allocating them based simply on peak demand. The resource provisioning problem will be posed as one of sequential decision making under uncertainty and solved using a predictive control scheme. We will also develop model-based diagnosis techniques that use the divergence between observed system behaviors and nominal model behaviors to isolate and characterize faults due to application errors and corrupted resources. Using a blade server cluster at Drexel University, we will verify the developed techniques for dynamic resource provisioning and fault diagnosis in a multi-tier e-commerce architecture hosted in a virtual server environment.
Friday, June 8th at 10 a.m.
Bossone 303
|
