Feedback Performance Control for Self-Managing Computer Systems: an LPV Control Theoretic Approach

Open Access
- Author:
- Qin, Wubi
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 17, 2007
- Committee Members:
- Dr Qian Wang, Committee Chair/Co-Chair
Asok Ray, Committee Member
Dr Kon Well Wang, Committee Member
Anand Sivasubramaniam, Committee Member - Keywords:
- linear parameter varying systems (lpv)
internet services
feedback control
performance management
quality of service (qos)
request level scheduling
web server - Abstract:
- High-performance server systems have been widely used in today's commercial and scientific server environments, for example, Web and Email applications, online trading and multimedia services, and many more. This research investigates the autonomic performance management of computer server systems in the context of Internet hosting centers. Two scenarios of the performance mangagement problem are studied: admission control and resource allocation. This research contributes to the problem in three major aspects: modeling of Internet services via new approaches, more realistic and advanced control designs, and intensive performance evaulations of various design solutions. Modeling wise, it is learned that the linear methods may not be adequate for Internet services modeling when there are large variations in load conditions. Consequently, more sophisticated nonlinear system modeling is needed to improve system performance and robustness with respect to large variations of load conditions. A linear uncertain model and a Linear-Parameter-Varying (LPV) model of hosting center server systems are derived based on theoretical analysis of transient queueing dynamics, as well as LPV system identification, where workload characterizing parameters are utilized as scheduling variables. Exisiting feedback control based performance management for Internet server systems relies on worst-case estimates of load and resource availability thus provisions resources to meet peak demands. Since the worst-case resource demand is likely to be significantly higher than its normal usage, these methods could be economically unfavorable. This work focuses on the development of an LPV control design for the performance management of server where time-varying effects of the system are explicitly considered by utilizing workload arrival and service parameters as scheduling variables. Further, it is widely noticed that workload parameters of Internet services are more or less unpredictable. This makes deterministic modeling and control design approaches either too conservative or unable to meet performance service level agreements (SLA). Probabilistic modeling and control design methods are proposed to balance between the risk of missing performance SLA and the resource efficiency. Request level approach to provide response time guarantee based on detailed analysis of requests response time components is studied. A request level scheduling algorithm is proposed that can be used with workload characterization and modeling techniques. Preliminary results in a response time guarantee problem show impressive request level performance over a wide range of sampling intervals. Intensive simulations are conducted using real Web server workloads. Performances of the proposed LPV based modeling and control design approaches are benchmarked using queueing theory based provisioning results and existing linear control modeling & designs. It is shown that the proposed method outperforms linear control and conventional queueing-theory based designs. These evaluations provide guidelines on which method to use for a particular problem. The results from this research can be easily generalized to accommodate more complicated models in characterizing workloads and server environments to enhance system performance.