Ent?te

Logo du LIFL

  1. News

Hdr defense of

Richard OLEJNIK

Monday 26 September 2011, 02h00 PM
Bâtiment des thèses

Scale up of Java distributed self-adaptive applications

Advisor : Bernard Toursel, Professor, Université Lille 1
Main examiners : Beniamino DI MARTINO, Professor, Second University of Naples (Italy)
Thierry PRIOL, Directeur de Recherche, INRIA Rennes - Bretagne Atlantique
Marek TUDRUJ, Professor, Sciences Academy of Poland
Members : Pierre BOULET, Professor, Université Lille 1
Christophe CERIN, Professor, Université Paris 13

A rapid evolution of networks, workstations, large supercomputers and personal computers, gives rise to new architectural alternatives for parallel and distributed computing. Clusters, grids and, more recently, cloud computing can therefore give an answer to constantly growing demands for computational resources, thanks to new paradigms, software concepts and systems, which are all based on distributed programming. The main features of distributed and heterogeneous applications are their irregularity and unpredictability. To enable efficient execution of such applications, we propose a programming environment for distributed applications in Java and the runtime environment ADAJ (Adaptive Distributed Applications in Java), which optimizes dynamic placement of application objects on clusters and computers within a grid. This distribution is based on a novel mechanism of observation of object work, and the relationships between the objects. The gain from this flexible and adaptive object distribution results in better execution efficiency and in better use of the power of different computers. At the same time it minimizes communication costs and reduces extra cost associated with application control. With these mechanisms, ADAJ provides automatic and adaptive distribution of application elements across the job execution platform, giving thus a reply to the computing and resource availability changes. This operation is based on a cycle stealing method and can control of the application execution granularity. As a result, the programmer does not need to worry about it anymore.

Mechanisms have been implemented for various platforms and technologies. Initially, they have been designed to run on clusters of workstations containing no more than a hundred computers. In order to scale up the solution designed for cluster computing, we have re-engineered, processed and completed it. Specifically, we have introduced a framework based on software components, to help the designer to build applications for grids of computers. This work was then extended so that the platform ADAJ is today a full middleware stack. It is based on web services and it’s information system is based on agents. Mechanisms of ADAJ can now manage grid execution platforms, consisting of thousands of nodes. Finally we have tested this approach on datamining problems with some distributed algorithms, which have been specifically developed. By this work, we have responded to the current problems concerning the implementation and use of grids by designing a new SOKU (Service Oriented Knowledge Utilities) architecture. Finally, we show how this research can be integrated in the theme of embedded systems.

Keywords: Grid computing, auto-adaptive applications, dynamic load balancing, SOKU systems, datamining
 

Calendar

Defenses

Ours

UMR 8022 - Laboratoire d'Informatique Fondamentale de Lille - Copyright © 2012 Sophie TISON - Crédits & Mentions légales

Page respectant XHTML et CSS.