|
Mianyu Wang
Advisor: Moshe Kam, Ph.D.
Abstract:
Web-based services such as online banking and retail commerce are often hosted on distributed computing systems comprising heterogeneous and networked servers. To operate such systems efficiently while satisfying stringent quality-of-service (QoS) requirements, multiple performance-related parameters must adapt to changing operating conditions. The workload a system must process might be time-varying, and hardware or software resources could fail or need to be replaced during system operation. To cope with their growing scale and complexity, such systems must become largely autonomic, namely be capable of managing with minimal human intervention.
In this study, we propose to develop a decentralized performance management framework using theories and techniques from model predictive control, optimal control, and hybrid dynamical systems. This framework will be used to manage computing systems that exhibit non-linear behavior, and whose performance must be optimized under various operating constraints. We first model a web-server cluster with one service level and multiple heterogeneous servers by a non-linear discrete-time state equation, and formulate the corresponding QoS control and power management problem as a multistage optimal control problem with constraints on a decentralized server node. We then model the cluster with multiple service levels as a class of discrete-time hybrid dynamical system, denoted as discrete hybrid automata (DHA). The combined QoS control, resource provisioning, and power management problem will be cast into a multistage optimal control problem of a hybrid dynamical system on a decentralized server node. We also propose to design a model predictive control based feedback strategy to implement the online decentralized controllers for autonomic distributed web systems. The framework and developed techniques will be verified through a discrete event simulator with real world workload data.
Monday, July 23rd at 3 p.m.
Bossone 303
|
