Modeling, Online Performance Monitoring, and Accelerated Evaluation of Network Services
Prof. Michael Devetsikiotis
NC State University
Wed, Mar 28, 12 Noon in McGl 020
Modeling and adaptation of resources based on state and workload (current or predicted) is highly desirable in emerging high-performance computing and information service systems, on the path towards completely "autonomic" services. In this seminar, we provide an overview of our efforts at NC State, in collaboration with IBM and Tekelec, to develop frameworks and algorithms for modeling of emerging network-based services, predictive and dynamic resource allocation, adaptive scheduling, on line performance monitoring and accelerated estimation.
We describe techniques in a "smart" control framework that includes quality of service and economic considerations. We present an overview of adaptive scheduling approaches that we have used for profit or utility-oriented scheduling in service access nodes. We are currently working to apply similar techniques to Web services, network appliances and multimedia services (e.g., SIP).
The ability to diagnose important events and dynamically adapt to transient conditions is essential to meet the end-user demands of an autonomic information system. On-line performance monitoring and sampling techniques, which analyze the state of the system while imposing a minimal footprint on available resources, are essential for on-line control algorithms in information and software systems.
Finally, we discuss related modeling and simulation techniques, including meta-modeling and fast simulation techniques based on Importance Sampling for accelerated estimation of extremely low probabilities in highly available services.
