AI Lab logo
menu MENU

Computer Engineering Seminar

Decentralized Control and Optimization Techniques for Autonomic Performance Management of Distributed Computing Systems

Nagarajan Kandasamy
SHARE:

A distributed computing system hosting e-commerce, business, and scientific applications must typically satisfy stringent quality-of-service (QoS) requirements while operating in a dynamic environment. For example, the workload to be processed may be time varying, and hardware and software components may fail during operation. To achieve QoS goals in such systems, numerous performance-related parameters must be continuously optimized to respond rapidly to both time-varying computing demands as well as component failures. As these systems increase in size and complexity, manually tuning key operating parameters will become very difficult. Therefore, it is highly desirable that future systems manage themselves, given only high-level guidance by administrators. Such autonomic computing systems aim to maintain the specified QoS by adaptively tuning key operating parameters with minimum manual intervention. This talk discusses how to apply concepts from model predictive and optimal control theory to provide the theoretical basis for enforcing self-managing behavior in computing systems – this is in contrast to current approaches that are largely heuristic and ad hoc. We will describe decentralized control architectures for autonomic performance management in a distributed system comprising multiple interacting components. As case studies, hierarchical and fully distributed control schemes are developed to manage the power consumption of a heterogeneous computer cluster while satisfying the QoS requirements of a time-varying workload. Using workload traces from Soccer World Cup 98, we show that the proposed methods are scalable, have low run-time overhead, and adapt quickly to time-varying workload patterns. Finally, we will discuss other uses for decentralized control including dynamic resource provisioning in utility computing and self-managing data streaming applications.

Nagarajan Kandasamy is an Assistant Professor with the Electrical and Computer Engineering Department at Drexel University. He received the PhD degree from the University of Michigan in 2003 where he was a research assistant at the Advanced Computer Architecture laboratory. Prior to joining Drexel, he was a research scientist at the Institute for Software Integrated Systems, Vanderbilt University. His interests include dependable computing, embedded systems, self-managing and distributed systems, and testing and verification of digital systems.

Sponsored by

ACAL