Die immer dichtere und schnellere Vernetzung von Rechnern und Rechenzentren über Hochgeschwindigkeitsnetzwerke ermöglicht eine neue Art des wissenschaftlich verteilten Rechnens, bei der geographisch weit auseinanderliegende Rechenkapazitäten zu einer Gesamtheit zusammengefasst werden können. Dieser so entstehende virtuelle Superrechner, der selbst aus mehreren Grossrechnern besteht, kann dazu genutzt werden Probleme zu berechnen, für die die einzelnen Grossrechner zu klein sind. Die Probleme, die numerisch mit heutigen Rechenkapazitäten nicht lösbar sind, erstrecken sich durch sämtliche Gebiete der heutigen Wissenschaft, angefangen von Astrophysik, Molekülphysik, Bioinformatik, Meteorologie, bis hin zur Zahlentheorie und Fluiddynamik um nur einige Gebiete zu nennen.<br /> <br /> Je nach Art der Problemstellung und des Lösungsverfahrens gestalten sich solche "Meta-Berechnungen" mehr oder weniger schwierig. Allgemein kann man sagen, dass solche Berechnungen um so schwerer und auch um so uneffizienter werden, je mehr Kommunikation zwischen den einzelnen Prozessen (oder Prozessoren) herrscht. Dies ist dadurch begründet, dass die Bandbreiten bzw. Latenzzeiten zwischen zwei Prozessoren auf demselben Grossrechner oder Cluster um zwei bis vier Grössenordnungen höher bzw. niedriger liegen als zwischen Prozessoren, welche hunderte von Kilometern entfernt liegen.<br /> <br /> Dennoch bricht nunmehr eine Zeit an, in der es möglich ist Berechnungen auf solch virtuellen Supercomputern auch mit kommunikationsintensiven Programmen durchzuführen. Eine grosse Klasse von kommunikations- und berechnungsintensiven Programmen ist diejenige, die die Lösung von Differentialgleichungen mithilfe von finiten Differenzen zum Inhalt hat. Gerade diese Klasse von Programmen und deren Betrieb in einem virtuellen Superrechner wird in dieser vorliegenden Dissertation behandelt. Methoden zur effizienteren Durchführung von solch verteilten Berechnungen werden entwickelt, analysiert und implementiert. Der Schwerpunkt liegt darin vorhandene, klassische Parallelisierungsalgorithmen zu analysieren und so zu erweitern, dass sie vorhandene Informationen (z.B. verfügbar durch das Globus Toolkit) über Maschinen und Netzwerke zur effizienteren Parallelisierung nutzen. Soweit wir wissen werden solche Zusatzinformationen kaum in relevanten Programmen genutzt, da der Grossteil aller Parallelisierungsalgorithmen implizit für die Ausführung auf Grossrechnern oder Clustern entwickelt wurde. / In order to face the rapidly increasing need for computational resources of various scientific and engineering applications one has to think of new ways to make more efficient use of the worlds current computational resources. In this respect, the growing speed of wide area networks made a new kind of distributed computing possible: Metacomputing or (distributed) Grid computing. This is a rather new and uncharted field in computational science. The rapidly increasing speed of networks even outperforms the average increase of processor speed: Processor speeds double on average each 18 month whereas network bandwidths double every 9 months. Due to this development of local and wide area networks Grid computing will certainly play a key role in the future of parallel computing.<br /> <br /> This type of distributed computing, however, distinguishes from the traditional parallel computing in many ways since it has to deal with many problems not occurring in classical parallel computing. Those problems are for example heterogeneity, authentication and slow networks to mention only a few. Some of those problems, e.g. the allocation of distributed resources along with the providing of information about these resources to the application have been already attacked by the Globus software.<br /> <br /> Unfortunately, as far as we know, hardly any application or middle-ware software takes advantage of this information, since most parallelizing algorithms for finite differencing codes are implicitly designed for single supercomputer or cluster execution. We show that although it is possible to apply classical parallelizing algorithms in a Grid environment, in most cases the observed efficiency of the executed code is very poor.<br /> <br /> In this work we are closing this gap. In our thesis, we will<br /> - show that an execution of classical parallel codes in Grid environments is possible but very slow<br /> - analyze this situation of bad performance, nail down bottlenecks in communication, remove unnecessary overhead and other reasons for low performance<br /> - develop new and advanced algorithms for parallelisation that are aware of a Grid environment in order to generelize the traditional parallelization schemes<br /> - implement and test these new methods, replace and compare with the classical ones - introduce dynamic strategies that automatically adapt the running code to the nature of the underlying Grid environment.<br /> <br /> The higher the performance one can achieve for a single application by manual tuning for a Grid environment, the lower the chance that those changes are widely applicable to other programs. In our analysis as well as in our implementation we tried to keep the balance between high performance and generality. None of our changes directly affect code on the application level which makes our algorithms applicable to a whole class of real world applications.<br /> <br /> The implementation of our work is done within the Cactus framework using the Globus toolkit, since we think that these are the most reliable and advanced programming frameworks for supporting computations in Grid environments. On the other hand, however, we tried to be as general as possible, i.e. all methods and algorithms discussed in this thesis are independent of Cactus or Globus.

Vernetzte Standardarbeitsplatzrechner (sog. Cluster) sind eine attraktive Umgebung zur Ausf"uhrung paralleler Programme; f"ur einige Anwendungsgebiete bestehen jedoch noch immer ungel"oste Probleme. Ein solches Problem ist die Verl"asslichkeit und Rechtzeitigkeit der Programmausf"uhrung: In vielen Anwendungen ist es wichtig, sich auf die rechtzeitige Fertigstellung eines Programms verlassen zu k"onnen. Mechanismen zur Kombination dieser Eigenschaften f"ur parallele Programme in verteilten Rechenumgebungen sind das Hauptanliegen dieser Arbeit. Zur Behandlung dieses Anliegens ist eine gemeinsame Metrik f"ur Verl"asslichkeit und Rechtzeitigkeit notwendig. Eine solche Metrik ist die Responsivit"at, die f"ur die Bed"urfnisse dieser Arbeit verfeinert wird. Als Fallstudie werden Calypso und Charlotte, zwei Systeme zur parallelen Programmierung, im Hinblick auf Responsivit"at untersucht und auf mehreren Abstraktionsebenen werden Ansatzpunkte zur Verbesserung ihrer Responsivit"at identifiziert. L"osungen f"ur diese Ansatzpunkte werden zu allgemeineren Mechanismen f"ur (parallele) responsive Dienste erweitert. Im Einzelnen handelt es sich um 1. eine Analyse der Responsivit"at von Calypsos ``eager scheduling'' (ein Verfahren zur Lastbalancierung und Fehlermaskierung), 2. die Behebung eines ``single point of failure,'' zum einen durch eine Responsivit"atsanalyse von Checkpointing, zum anderen durch ein auf Standardschnittstellen basierendes System zur Replikation bestehender Software, 3. ein Verfahren zur garantierten Ressourcenzuteilung f"ur parallele Programme und 4.die Einbeziehung semantischer Information "uber das Kommunikationsmuster eines Programms in dessen Ausf"uhrung zur Verbesserung der Leistungsf"ahigkeit. Die vorgeschlagenen Mechanismen sind kombinierbar und f"ur den Einsatz in Standardsystemen geeignet. Analyse und Experimente zeigen, dass diese Mechanismen die Responsivit"at passender Anwendungen verbessern. / Clusters of standard workstations have been shown to be an attractive environment for parallel computing. However, there remain unsolved problems to make them suitable to some application scenarios. One of these problems is a dependable and timely program execution: There are many applications in which a program should be successfully completed at a predictable point of time. Mechanisms to combine the properties of both dependable and timely execution of parallel programs in distributed computing environments are the main objective of this dissertation. Addressing these properties requires a joint metric for dependability and timeliness. Responsiveness is such a metric; it is refined for the purposes of this work. As a case study, Calypso and Charlotte, two parallel programming systems, are analyzed and their shortcomings on several abstraction levels with regard to responsiveness are identified. Solutions for them are presented and generalized, resulting in widely applicable mechanisms for (parallel) responsive services. Specifically, these solutions are: 1) a responsiveness analysis of Calypso's eager scheduling (a mechanism for load balancing and fault masking), 2) ameliorating a single point of failure by a responsiveness analysis of checkpointing and by a standard interface-based system for replication of legacy software, 3) managing resources in a way suitable for parallel programs, and 4) using semantical information about the communication pattern of a program to improve its performance. All proposed mechanisms can be combined and are suitable for use in standard environments. It is shown by analysis and experiments that these mechanisms improve the responsiveness of eligible applications.

Three-dimensional volume data is routinely produced, at increasingly high spatial resolution, in computer simulations and image acquisition tasks. In-situ visualization, the visualization of an experiment or simulation while it is running, enables new modes of interaction, including simulation steering and experiment control. These can provide the scientist a deeper understanding of the underlying phenomena, but require interactive visualization with smooth viewpoint changes and zooming to convey depth perception and spatial understanding. As the size of the volume data increases, however, it is increasingly challenging to achieve interactive visualization with smooth viewpoint changes. This thesis presents an end-to-end solution for interactive in-situ visualization based on novel extensions proposed to the Volumetric Depth Image (VDI) representation. VDIs are view-dependent, compact representations of volume data than can be rendered faster than the original data. Novel methods are proposed in this thesis for generating VDIs on large data and for rendering them faster. Together, they enable interactive in situ visualization with smooth viewpoint changes and zooming for large volume data. The generation of VDIs involves decomposing the volume rendering integral along rays into segments that store composited color and opacity, forming a representation much smaller than the volume data. This thesis introduces a technique to automatically determine the sensitivity parameter that governs the decomposition of rays, eliminating the need for manual parameter tuning in the generation of a VDI. Further, a method is proposed for sort-last parallel generation and compositing of VDIs on distributed computers, enabling their in situ generation with distributed numerical simulations. A low latency architecture is proposed for the sharing of data and hardware resources with a running simulation. The resulting VDI can be streamed for interactive visualization. A novel raycasting method is proposed for rendering VDIs. Properties of perspective projection are exploited to simplify the intersection of rays with the view-dependent segments contained within the VDI. Spatial smoothness in volume data is leveraged to minimize memory accesses. Benchmarks are performed showing that the method significantly outperforms existing methods for rendering the VDI, and achieves responsive frame rates for High Definition (HD) display resolutions near the viewpoint of generation. Further, a method is proposed to subsample the VDI for preview rendering, maintaining high frame rates even for large viewpoint deviations. The quality and performance of the approach are analyzed on multiple datasets, and the contributions are provided as extensions of established open-source tools. The thesis concludes with a discussion on the strengths, limitations, and future directions for the proposed approach.

No description available.