Performance Modeling and Resource Allocation for Adaptive Agent-Based Systems
Open Access
- Graduate Program:
- Industrial Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- November 01, 2007
- Committee Members:
- Soundar Rajan Tirupatikumara, Committee Chair
- M Jeya Chandra, Committee Member
- Tao Yao, Committee Member
- Dr Jun Shu, Committee Member
- Dr Natarajan Gautam, Committee Member
- Keywords:
- queueing
- performance
- resource allocation
- agents
- logistics
- supply-chain
- optimization
- Abstract:
- Applications are increasingly becoming networks of individual application components spread over an infrastructure of physical resources (servers and computing entities). Such distributed agent-based applications are not only prevalent in military domains but also in commercial domains such as data centers. To assure the survivability of a distributed application, repeatedly estimating what sort of multi-dimensional guarantees (such as response time or availability) can be made as a function of the offered load and environmental conditions is paramount. This thesis studies how such guarantees can be computed and negotiated in a distributed MAS that is situated on an unreliable infrastructure. Firstly, this work concentrates on identifying relevant service-level attributes that can be estimated for the failure-prone infrastructure. Queueing models with single and multi-class traffic are studied in scenarios with breakdown, repair and catastrophic failure, and are utilized for the estimation of performance and reliability attributes in steady state. These analytical results serve as internal models for the agents which aid them in evaluating the quality of service that can be attained given the environmental conditions (stresses and information load). Secondly, two resource allocation mechanisms are studied that are used to assign the agents to the nodes forming the infrastructure which are constrained by capacity, and are non-homogeneous in terms of their capability and the stresses experienced. Since the assignment problem is in general NP-hard, a few allocation heuristics with varying roles on the part of the agents and the infrastructure are proposed. Finally, in order to use resources efficiently, a usage price is computed based on the application’s demand for the service-level attributes (which are the information goods sold by the infrastructure). The main contributions of this work are the analytical solutions for the queueing models – which include multi-class traffic, service disruption (both temporary and catastrophic failure) and non-preemptive priority scheduling. The analytical models pave the way for rapid negotiation between the MAS and the infrastructure as opposed to relatively slower simulation models. The analytical solutions are applicable to other domains including web-servers and hosting. The other contribution is the study of decentralized mechanisms that aid in the negotiation of quantitative service-level agreements in multi-agent systems and, in general, service-oriented architectures. While negotiation protocols and methods are widely studied, automatic negotiation using internal models are novel to this work, especially to distributed MAS.