DYNAMIC LOAD BALANCING FOR ACTIVE OBJECTS ON COMPUTER GRIDS

Javier Bustos-Jiménez 1
1 OASIS - Active objects, semantics, Internet and security
CRISAM - Inria Sophia Antipolis - Méditerranée , COMRED - COMmunications, Réseaux, systèmes Embarqués et Distribués
Abstract : This thesis aims to set the foundations for the development of load-balancing algorithms for the
active objects model defined by ProActive in the context of large-scale networks (Grids).
ProActive is an open-source Java middleware which achieves seamless programming for concurrent,
parallel, distributed, and mobile computing, implementing the active-object paradigm. In
ProActive, each active object has its own control thread and can independently decide in which
order to serve incoming method calls. Incoming method calls are automatically stored in a queue
of pending requests (called a service queue). To add efficiency to the active objects paradigm,
ProActive provides a migration mechanism, having automatic location and transparency through
the use of forwarders. The migration operation comes with a communication penalty: an active
object must migrate with its complete state, consisting of its pending requests (method calls), futures,
and passive (mandatory non-shared) objects. Therefore, ProActive applications are sensitive
to latency.
When several active objects with identical functionality are deployed, a load-balancing algorithm
can be used to improve the performance of an application using that functionality. We call
workload the total work units required by an application to finish. To use some active object's functionality,
an application puts a work unit into the service queue of that active object. The workload
can be balanced across several active objects either by sending active objects from a highly loaded
processor to a less loaded one, or by stealing active objects from a highly loaded processor by a
less loaded one. The environment where the active objects run is usually composed of multiple
clusters of resources, such as sets of machines interconnected by a high-speed local network.
In this thesis, a load-balancing algorithm for active objects which belong to a parallel application
was developed and studied, setting the foundations for development of load-balancing
algorithms for the middleware ProActive. This first approach is called the Robin-Hood + Nottingham
Sheriff load-balancing algorithm. This algorithm was validated in the context of a large-scale
network (over 1,000 nodes) through simulations. Therefore, we presented our model of Grids
based on observation and measurement of what we consider key characteristics for active-objects
load-balancing: processing capacity and inter-resource communication latency.
Finally, we present coupling contracts for deployment of parallel applications, showing how to
use them in the context of load balancing by now choosing between a local scheduler (for clusters)
and ProActive's load balancer.
Type de document :
Thèse
Networking and Internet Architecture [cs.NI]. Université Nice Sophia Antipolis, 2006. English
Liste complète des métadonnées

Littérature citée [121 références]  Voir  Masquer  Télécharger

https://tel.archives-ouvertes.fr/tel-00164582
Contributeur : Javier Bustos-Jiménez <>
Soumis le : vendredi 20 juillet 2007 - 18:47:16
Dernière modification le : mardi 22 mars 2016 - 01:26:47
Document(s) archivé(s) le : lundi 24 septembre 2012 - 11:20:33

Fichier

Identifiants

  • HAL Id : tel-00164582, version 1

Collections

Citation

Javier Bustos-Jiménez. DYNAMIC LOAD BALANCING FOR ACTIVE OBJECTS ON COMPUTER GRIDS. Networking and Internet Architecture [cs.NI]. Université Nice Sophia Antipolis, 2006. English. 〈tel-00164582〉

Partager

Métriques

Consultations de
la notice

854

Téléchargements du document

665