PERFORMANCE AND CONTROL OF PARALLEL MULTI-SERVER QUEUES WITH APPLICATIONS to WEB HOSTING Services

Open Access
Author:
Chen, Yiyu
Graduate Program:
Industrial Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
June 01, 2006
Committee Members:
  • Soundar Rajan Tirupatikumara, Committee Chair
  • Natarajan Gautam, Committee Chair
  • Anand Sivasubramaniam, Committee Member
  • Dennis Kon Jin Lin, Committee Member
Keywords:
  • power management
  • multi-server queueing analysis
  • QoS
  • web hosting center
Abstract:
This thesis addresses power and performance management of a queueing system as follows. It consists of multiple parallel queues, each of which serves single type of requests and is called an application. For each queue, there are possibly multiple servers, with some scheduling discipline and waiting area having infinite capacity. Each request brings a random amount of work and will be assigned by application scheduler to a server and processed at a particular speed. Web hosting center is an example of such a system. Therein operates thousands of servers consuming huge amount of energy and serves hundreds of applications that require the system to deliver Quality of Service (QoS). The complexity of managing such huge systems in cost effective manners is not a trivial endeavor, especially under stringent QoS requirements from self-similar like traffic. This thesis proposes a three-tiered solution methodology at system, application, and server level as a tunable resource allocation strategy for power and performance management (for such aforementioned systems). At system level, a centralized G/G/m-based solving algorithm, applicable at different time-granularities, is developed to provision servers at required operating frequencies. A packing/scheduling mechanism is built for dispatching incoming workload into a given set of servers to balance energy-savings and performance needs in the second tier. Further, in the third tier, a decentralized frequency control strategy for an individual server is devised by adopting fluid modeling and Markov Decision Process(MDP). In terms of contributions, this thesis provides an energy aware resource management framework for web hosting centers. The proposed three-tired solution methodology can achieve significantly high amount of energy-savings and maintain the system performances as well, compared to those done by prior studies. Secondly, it is one of the first few studies, which provides a rigorous optimization formulation under self-similar-like workloads with respect to availability of controllable resources. Solutions derived from those optimization can be easily computed and implemented in an online fashion for computer systems. Finally, the proposed solution methodology can be generalized as a strategy of resource management for any parallel multi-server queueing system serving multiple classes of requests, such as call centers.