Dynamische Lastverteilung für middlewarebasierte Softwaresysteme in heterogenen Rechnerumgebungen

The rising complexity of algorithms and conceptual improvements in the field of software engineering causes an increasing demand for computing power, despite of the technological progress in hardware development. Especially scientific and engineering applications, for example weather and climate calculations, but also image processing and artificial intelligence generate high computational loads. In this context, parallel computers, multi processor computers and an increasing amount of networked computers like Clusters of Workstations (COW), Grids or Networks of Workstations (NOW) are used to satisfy the demands. While cluster is synonymous for a group of homogeneous workstations in an isolated, narrow environment, a network of workstations describes a collection of mostly heterogeneous computers, interconnected within local area networks, for example in an office or laboratory environment. Another aspect in such networks is the occurrence of concurrent users. In this application scenario a common method for the interaction of software components is provided by middleware, a software layer that allows communication between components on a very high abstraction level. These aspects aggravate the efficient usage of available hardware resources and have rarely been investigated in the context of middleware-based load balancing. The subject of this thesis is the development of dynamic load distribution and balancing techniques in the middleware layer for heterogeneous computing systems to increase the performance of interactive, resource intensive applications. The thesis considers object oriented and component based programming concepts, which have become very important during the last years. In order to support these paradigms, the communication protocols chosen in this thesis provide the corresponding abstraction layers. The thesis presents a concept as well as a mathematical system model for dynamic load balancing in the Common Object Request Broker Architecture (CORBA). The proposed solution is the first one to take capacity parameters of heterogeneous networked workstations into consideration. Based on this model, a simulation is developed to evolve and optimize a dynamic load balancing algorithm. The execution time of previously distributed calculations is determined and used to prevent improper load migrations. The developed concept and its algorithms allow load balancing based on current system states, such as cpu-clockrates, memory consumption and network parameters. It distributes workload more efficiently to available computing resources than previous standard algorithms in the middleware layer and decreases execution times. The comparison with state of the art technology in the simulation, with benchmarks and within an example application from the field of artificial intelligence and image processing demonstrates the benefits of the developed architecture.

Grantor dc:publisher

Publikationsserver der RWTH Aachen University

Year dc:date

2004

Author dc:creator

• Krüger, Thomas Moritz

Contributors dc:contributor

• Kraiss, Karl-Friedrich

• info:eu-repo/classification/ddc/004
• Hochleistungsrechnen
• Heterogenes Rechnernetz
• Middleware
• Dynamische Lastteilung
• CORBA
• Benchmark
• Informatik
• dynamic load balancing
• load distribution

Statement dc:rights

• info:eu-repo/semantics/openAccess

Language dc:language

ger

Identifier

https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-121281%22

OAI identifier oai:identifier

oai:publications.rwth-aachen.de:59502