Disputas: Sharath Babu Musunoori

Prøveforelesning

Bedømmelseskomité

Professor B. John Oommen, School of Computer Science, Carleton University, Ottawa, Canada
Professor Geoff Coulson, Computing Department, Faculty of Applied Sciences, Lancaster University, UK
Associate Professor Carsten Griwodz, Institutt for informatikk, Universitetet i Oslo

Leder av disputas:  Morten Dæhlen

Veileder:  Frank Eliassen, Richard Staehli

Sammendrag

This is a doctoral thesis submitted to the Department of Informatics, Faculty of Mathematics and Natural Scineces, University of Oslo in partial fulfillment for the degree of Doctor of Philosophy in Computer Science for the year 2007. The work reported in this thesis has been carried out within the context of the QoS-Aware Component Architecture (QuA) project at the Networks and Distributed Systems department, Simula Research Laboratory.

The most successful existing distributed physical (e.g. electrical power, railroads, and telephone) infrastructures today have proved their importance to mankind. Similarly, continuous improvements in computer hardware, software and significant research results in the science and engineering areas have led to the development of large computational platforms, called GRIDs.

Computational grids are widely accepted as future platforms for parallel and distributed systems, since grid middleware aims to solve problems of resource sharing and management across organizational domains. Apparently, there is a need to balance the computational resources of the grid to minimize the overall costs. But at the same time, any application running on a computational grid, must get the necessary resources without knowing how these resources are arranged. In particular, achieving acceptable application performance in a computational grid environment remains a difficult engineering challenge.

The fundamental problem addressed in this thesis is the configuration of an application for a dynamic grid environment. This is the problem of stochastically partitioning the application's service set onto a set of heterogeneous capsules (computing machines), each of which offers a set of resources to the running services such that no capsule is oversubscribed with respect to its resources and all the services achieve the minimum required quality necessary for an application to run as intended.

This thesis introduces a set of partitioning algorithms that are developed on the basis of the basic principles of learning automata, ant system and their combination. A learning automaton (LA) basically proposes an action to a stochastic environment that returns binary feedback (a reward or a penalty). Through several such iterations the automaton will learn which action is most likely to give a positive feedback. On the other hand, the use of ant system concepts in this work was inspired by the natural behavior of social elements (a colony of ants or wasps) of an ecosystem. The foraging behavior of multiple ants creates a collective intelligence, which in nature is capable of generating useful solutions. In addition to the use of the metaphor of foraging unintelligent ants of an ant colony, the basic algorithms are extended with the concept of learning automata, thus making the ants and other working agents more intelligent.

The algorithms were tested and the results analyzed through simulations on randomly generated sets of application services and grid environments. The results obtained from simulations show that the proposed methods provide viable solutions to the complex problem of application service partitioning in grid-like environments. The work presented in this thesis demonstrates that learning and nature inspired approaches are capable of working with distributed decisions in random distributed environments. Additionally, the application models, simulation results and the detailed analysis of the proposed application service configuration algorithms are important contributions towards the vision of grid computing.

Kontaktperson

For mer informasjon, kontakt Lena Korsnes.