Feedback based performance management and fault tolerance for networked and embedded computing systems
Lui Sha, Xue Liu · 2006
Performance management and fault tolerance are two important issues faced by computing systems research. In this dissertation, we exploit the use of feedback control for performance management and fault tolerance. Specifically, we propose Queueing Model Based Feedback Control scheme to achieve performance regulation. Traditionally, queueing theory was used for modeling computing system's performance. It usually serves as an offline capacity planning tool. On the other hand, feedback control theory was used for dynamically controlling the performance of electro-mechanical systems. How to utilize the descriptive power of queueing theory and the prescriptive power of feedback control to control computing system's performance was an open problem. Queueing Model Based Feedback Control answers this question by integrating the strengths of both queueing model and feedback control into one unified framework. It provides better performance regulation compared with other control-theoretic approaches. We show the advantages of Queueing Model Based Feedback Control for two networked server applications: one is response time regulation for Apache Web server via dynamic resource allocations; the other is response time regulation for a multi-tiered Web service application via dynamic admission control. In the second part of this dissertation, we further exploit the use of feedback control to achieve fault tolerance for real-time embedded control systems. We propose ORTGA (On-demand Real-Time GuArd), a new fault tolerance architecture which utilizes feedback control based software execution. ORTGA delivers the same functionalities as previously proposed Simplex architecture, with the same high fault coverage and reliability but with much more efficient resource utilization and flexibility. Hence it can be deployed in a wide range of real-time embedded applications to provide fault tolerance. We implemented ORTGA in an inverted pendulum testbed to demonstrate its efficacy and efficiency. Based on the ORTGA design, we discussed the fault tolerance and scheduling co-design problem and its solutions.