Universität Stuttgart
Permanent URI for this communityhttps://elib.uni-stuttgart.de/handle/11682/1
Browse
Search Results
Item Open Access A massively parallel combination technique for the solution of high-dimensional PDEs(2018) Heene, Mario; Pflüger, Dirk (Jun.-Prof. Dr.)The solution of high-dimensional problems, especially high-dimensional partial differential equations (PDEs) that require the joint discretization of more than the usual three spatial dimensions and time, is one of the grand challenges in high performance computing (HPC). Due to the exponential growth of the number of unknowns - the so-called curse of dimensionality, it is in many cases not feasible to resolve the simulation domain as fine as required by the physical problem. Although the upcoming generation of exascale HPC systems theoretically provides the computational power to handle simulations that are out of reach today, it is expected that this is only achievable with new numerical algorithms that are able to efficiently exploit the massive parallelism of these systems. The sparse grid combination technique is a numerical scheme where the problem (e.g., a high-dimensional PDE) is solved on different coarse and anisotropic computational grids (so-called component grids), which are then combined to approximate the solution with a much higher target resolution than any of the individual component grids. This way, the total number of unknowns being computed is drastically reduced compared to the case when the problem is directly solved on a regular grid with the target resolution. Thus, the curse of dimensionality is mitigated. The combination technique is a promising approach to solve high-dimensional problems on future exascale systems. It offers two levels of parallelism: the component grids can be computed in parallel, independently and asynchronously of each other; and the computation of each component grid can be parallelized as well. This reduces the demand for global communication and synchronization, which is expected to be one of the limiting factors for classical discretization techniques to achieve scalability on exascale systems. Furthermore, the combination technique enables novel approaches to deal with the increasing fault rates expected from these systems. With the fault-tolerant combination technique it is possible to recover from failures without time-consuming checkpoint-restart mechanisms. In this work, new algorithms and data structures are presented that enable a massively parallel and fault-tolerant combination technique for time-dependent PDEs on large-scale HPC systems. The scalability of these algorithms is demonstrated on up to 180225 processor cores on the supercomputer Hazel Hen. Furthermore, the parallel combination technique is applied to gyrokinetic simulations in GENE, a software for the simulation of plasma microturbulence in fusion devices.Item Open Access Coupling schemes and inexact Newton for multi-physics and coupled optimization problems(2018) Scheufele, Klaudius; Mehl, Miriam (Prof. Dr.)This work targets mathematical solutions and software for complex numerical simulation and optimization problems. Characteristics are the combination of different models and software modules and the need for massively parallel execution on supercomputers. We consider two different types of multi-component problems in Part I and Part II of the thesis: (i) Surface coupled fluid- structure interactions and (ii) analysis of medical MR imaging data of brain tumor patients. In (i), we establish highly accurate simulations by combining different aspects such as fluid flow and arterial wall deformation in hemodynamics simulations or fluid flow, heat transfer and mechanical stresses in cooling systems. For (ii), we focus on (a) facilitating the transfer of information such as functional brain regions from a statistical healthy atlas brain to the individual patient brain (which is topologically different due to the tumor), and (b) to allow for patient specific tumor progression simulations based on the estimation of biophysical parameters via inverse tumor growth simulation (given a single snapshot in time, only). Applications and specific characteristics of both problems are very distinct, yet both are hallmarked by strong inter-component relations and result in formidable, very large, coupled systems of partial differential equations. Part I targets robust and efficient quasi-Newton methods for black-box surface-coupling of parti- tioned fluid-structure interaction simulations. The partitioned approach allows for great flexibility and exchangeable of sub-components. However, breaking up multi-physics into single components requires advanced coupling strategies to ensure correct inter-component relations and effectively tackle instabilities. Due to the black-box paradigm, solver internals are hidden and information exchange is reduced to input/output relations. We develop advanced quasi-Newton methods that effectively establish the equation coupling of two (or more) solvers based on solving a non-linear fixed-point equation at the interface. Established state of the art methods fall short by either requiring costly tuning of problem dependent parameters, or becoming infeasible for large scale problems. In developing parameter-free, linear-complexity alternatives, we lift the robustness and parallel scalability of quasi-Newton methods for partitioned surface-coupled multi-physics simulations to a new level. The developed methods are implemented in the parallel, general purpose coupling tool preCICE. Part II targets MR image analysis of glioblastoma multiforme pathologies and patient specific simulation of brain tumor progression. We apply a joint medical image registration and biophysical inversion strategy, targeting at facilitating diagnosis, aiding and supporting surgical planning, and improving the efficacy of brain tumor therapy. We propose two problem formulations and decompose the resulting large-scale, highly non-linear and non-convex PDE-constrained optimization problem into two tightly coupled problems: inverse tumor simulation and medical image registration. We deduce a novel, modular Picard iteration-type solution strategy. We are the first to successfully solve the inverse tumor-growth problem based on a single patient snapshot with a gradient-based approach. We present the joint inversion framework SIBIA, which scales to very high image resolutions and parallel execution on tens of thousands of cores. We apply our methodology to synthetic and actual clinical data sets and achieve excellent normal-to-abnormal registration quality and present a proof of concept for a very promising strategy to obtain clinically relevant biophysical information. Advanced inexact-Newton methods are an essential tool for both parts. We connect the two parts by pointing out commonalities and differences of variants used in the two communities in unified notation.Item Open Access Advancing manipulation skill learning towards sample-efficiency and generalization(2018) Englert, Peter; Toussaint, Marc (Prof. Dr.)Item Open Access Replicated execution of workflows(2018) Schäfer, David Richard; Rothermel, Kurt (Prof. Dr. rer. nat. Dr. h. c.)Workflows are the de facto standard for managing and optimizing business processes. Workflows allow businesses to automate interactions between business locations and partners residing anywhere on the planet. This, however, requires the workflows to be executed in a distributed and dynamic environment, where device and communication failures occur quite frequently. In case that a workflow execution becomes unavailable through such failures, the business operations that rely on the workflow might be hindered or even stopped, implying the loss of money. Consequently, availability is a key concern when using workflows in dynamic environments. In this thesis, we propose replication schemes for workflow engines to ensure the availability of the workflows that are executed by these engines. Of course, a workflow that is executed by a replicated workflow engine has to yield the same result as a non-replicated execution of that workflow. To this end, we formally define the equivalence of a replicated and a non-replicated execution called Single-Execution-Equivalence. Subsequently, we present replication schemes for both imperative and declarative workflow languages. Imperative workflow languages, such as the Web Service Business Process Execution Language (WS-BPEL), specify the execution order of activities through an ordering relation and are the predominant way of specifying workflow models. We implement a proof-of-concept for demonstrating the compatibility of our replication schemes with current (imperative) workflow technology. Declarative workflow languages provide greater flexibility by allowing the reordering of the activities within a workflow at run-time. We exploit this by executing differently ordered replicas on several nodes in the network for improving availability further.Item Open Access Learning structured models for active planning : beyond the Markov paradigm towards adaptable abstractions(2018) Lieck, Robert; Toussaint, Marc (Prof. Dr.)Item Open Access Window-based data parallelization in complex event processing(2018) Mayer, Ruben; Rothermel, Kurt (Prof. Dr. rer. nat. Dr. h. c.)With the proliferation of streaming data from sources such as sensors in the Internet of Things (IoT), situational aware applications become possible. Such applications react to situations in the surrounding world that are signaled by complex event patterns that occur in the sensor streams. In order to detect the patterns that correspond to the situations of interest, Complex Event Processing (CEP) is the paradigm of choice. In CEP, a distributed operator graph is spanned between the event sources and the applications. Each operator step-wise detects event patterns on subsequences, called windows, of its input stream and forwards output events that signal the detection to its successors. To cope with the ever-increasing workload at the operators, operator parallelization becomes necessary. To this end, data parallelization is a powerful paradigm, building on an architecture that consists of a splitter, operator instances and a merger, to scale up and scale out CEP operators. In doing so, the operators need to provide consistent output streams, i.e., not produce false-negatives or false-positives, keep a latency bound on pattern detection, elastically adapt their resource reservations to the workload, and be fault-tolerant against node and network failures. Related work has proposed data parallelization techniques that build on splitting the input event streams of an operator either in a key-based, a batch-based or a pane-based way. These approaches, however, only support a limited range of CEP operators. The goals of this thesis are (i) to support data parallelization for all window-based CEP operators, (ii) to develop adaptation methods such that CEP operators can keep a user-defined latency bound while minimizing costs for computing and networking resources, and (iii) to develop recovery methods that guarantee fault-tolerance at a low run-time overhead. To this end, the following contributions are made. First, we propose a window-based data parallelization method that is based on the externalization of the operator's window policy to a data parallelization framework. Second, basing on Queuing Theory, we propose a method to adapt the operator parallelization degree at run-time to the workload such that probabilistic bounds on the event buffering in the operator can be met. Third, we propose a batch scheduling algorithm that is able to assign subsequent overlapping windows to the same operator instance, so that communication overhead is minimized, while a latency bound in the operator instances is still kept. Forth, we propose a framework for parallel processing of inter-dependent windows that is based on the speculative processing of multiple versions of multiple windows in parallel. Fifth, we propose a lightweight rollback recovery method for CEP operator networks that exploits the externalization of the operator window policy to allow for the recovery of an arbitrary number of operators.Item Open Access Anforderungsbasierte Modellierung und Ausführung von Datenflussmodellen(2018) Hirmer, Pascal; Mitschang, Bernhard (Prof. Dr.)Heutzutage steigen die Menge an Daten sowie deren Heterogenität, Änderungshäufigkeit und Komplexität stark an. Dies wird häufig als das "Big-Data-Problem" bezeichnet. Durch das Aufkommen neuer Paradigmen, wie dem Internet der Dinge oder Industrie 4.0, nimmt dieser Trend zukünftig noch weiter zu. Die Verarbeitung, Analyse und Visualisierung von Daten kann einen hohen Mehrwert darstellen, beispielsweise durch die Erkennung bisher unbekannter Muster oder durch das Vorhersagen von Ereignissen. Jedoch stellen die Charakteristiken von Big-Data, insbesondere die große Datenmenge und deren schnelle Änderung, eine große Herausforderung für die Verarbeitung der Daten dar. Herkömmliche, bisher angewandte Techniken, wie zum Beispiel Analysen basierend auf relationalen Datenbanken, kommen hierbei oft an ihre Grenzen. Des Weiteren ändert sich auch die Art der Anwender der Datenverarbeitung, insbesondere in Unternehmen. Anstatt die Datenverarbeitung ausschließlich von Programmierexperten durchzuführen, wächst die Anwendergruppe auch um Domänennutzer, die starkes Interesse an Datenanalyseergebnissen haben, jedoch diese nicht technisch umsetzen können. Um die Unterstützung von Domänennutzern zu ermöglichen, entstand ca. im Jahr 2007, im Rahmen der Web-2.0-Bewegung, das Konzept der Mashups, die es auf einfachem Wege erlauben sollen, Anwender aus unterschiedlichen Domänen beim Zusammenführen von Programmen, grafischen Oberflächen, und auch Daten zu unterstützen. Hierbei lag der Fokus vor allem auf Webdatenquellen wie RSS-Feeds, HTML-Seiten, oder weiteren XML-basierten Formaten. Auch wenn die entstandenen Konzepte gute Ansätze liefern, um geringe Datenmengen schnell und explorativ durch Domänennutzer zu verarbeiten, können sie mit den oben genannten Herausforderungen von Big-Data nicht umgehen. Die Grundidee der Mashups dient als Inspiration dieser Dissertation und wird dahingehend erweitert, moderne, komplexe und datenintensive Datenverarbeitungs- und Analyseszenarien zu realisieren. Hierfür wird im Rahmen dieser Dissertation ein umfassendes Konzept entwickelt, das sowohl eine einfache Modellierung von Datenanalysen durch Domänenexperten ermöglicht - und somit den Nutzer in den Mittelpunkt stellt - als auch eine individualisierte, effiziente Ausführung von Datenanalysen und -verarbeitung ermöglicht. Unter einer Individualisierung wird dabei verstanden, dass die funktionalen und nichtfunktionalen Anforderungen, die je nach Anwendungsfall variieren können, bei der Ausführung berücksichtigt werden. Dies erfordert einen dynamischen Aufbau der Ausführungsumgebung. Hierbei wird dem beschriebenen Problem durch mehrere Ebenen begegnet: 1) Die Modellierungsebene, die als Schnittstelle zu den Domänennutzern dient und die es erlaubt Datenverarbeitungsszenarien abstrakt zu modellieren. 2) Die Modelltransformationsebene, auf der das abstrakte Modell auf verschiedene ausführbare Repräsentationen abgebildet werden kann. 3) Die Datenverarbeitungsebene, mit der die Daten effizient in einer verteilten Umgebung verarbeitet werden, und 4) die Datenhaltungsebene, in der Daten heterogener Quellen extrahiert sowie Datenverarbeitungs- oder Analyseergebnisse persistiert werden. Die Konzepte der Dissertation werden durch zugehörige Publikationen in Konferenzbeiträgen und Fachmagazinen gestützt und durch eine prototypische Implementierung validiert.