7. HPC-Status-Konferenz der Gauß-Allianz
Höchstleistungsrechenzentrum Stuttgart (HLRS), Stuttgart, 4. - 5. Dez. 2017

Das Forschungsprojekt SeASiTe (Selbstadaption für zeitschrittbasierte Simulationstechniken auf heterogenen HPC-Systemen) stellt sich der Aufgabe, eine systematische Untersuchung von Selbstadaption für zeitschrittbasierte Simulationstechniken auf heterogenen HPC-Systemen durchzuführen. Das Ziel ist der Entwurf und die Bereitstellung des Prototypen eines Werkzeugkastens, mit dessen Hilfe Programmierer ihre Anwendungen mit effizienten Selbstadaptionstechniken ausstatten können. Der Ansatz beinhaltet die Selbstadaption sowohl hinsichtlich relevanter System- und Programmparameter als auch möglicher Programmtransformationen. Die Optimierung der Programmausführung für mehrere nicht-funktionale Ziele (z.B. Laufzeit oder Energieverbrauch) soll auf einer Performance-Modellierung zur Eingrenzung des Suchraums effizienter Programmvarianten aufbauen. Anwendungsunabhängige Methoden und Strategien zur Selbstadaption sollen in einem Autotuning-Navigator gekapselt werden.

Phase-field models are an important class of mathematical techniques for the description of a multitude of industry-relevant physical and technical processes. The price for the broad applicability and mathematical elegance of this approach is the significant computing cost required for the simulation of phase-field equations at large scales. The significant complexity of the simulation demand the use of modern HPC architectures as well as novel parallelization techniques. In this talk, we will discuss the HPC aspects of the BMBF project “ParaPhase”, focussing on the parallelization strategies for spatial and temporal dimensions. Using multigrid techniques in both space and time, the consortium aims at providing efficient, flexible and high-order solvers for phase-field models. We will show first results for fracture propagation problems and discuss potential extensions and challenges.

Die Vorhersage der Ausbreitung von Wellen spielt in einer Vielzahl von Anwendungen eine wesentliche Rolle, ohne die die Zukunft in einer technologisch-wissenschaftlich orientierten Welt nicht vorstellbar wäre. Das Projekt WAVE entwickelt eine Toolbox mit Implementationen von FD-Verfahren. Im aktuellen Bericht wird der Fokus auf der Weiterentwicklung des quelloffenen Frameworks LAMA in diesem Zusammenhang liegen.

The direct numerical computation of acoustic noise (DNC) generated by fluid flows offers a number of advantages compared to a hybrid approach, in which both fields are solved independently. In DNC, hydrodynamics and acoustics are solved in a coupled manner which allows the detection and numerical representation of intricate interactions between both fields, e.g. feedback loops which generate tonal noise. However, this approach intrinsically requires the resolution of the disparate length and time-scales of non-linear turbulent production and the acoustic propagation, making DNC a challenging example of a multiscale problem. High order Discontinuous Galerkin (DG) methods have gained significant attention as baseline schemes for these types of problems. An open-source framework based on a computationally efficient variant of the DG method has been developed by the Numerics Research group and has been successfully applied to DNS and LES of turbulent flows. In particular, this framework has proven its capabilities for DNC in predicting sensitive acoustic feedback loops in complex geometries. However, due to the strong non-linear processes involved, the occurrence of acoustic feedback phenomena are highly sensitive to turbulence, geometry details and the quality of the numerical scheme and its generation is still an active field of research. Therefore, the aim of this project is to develop an uncertainty quantification (UQ) framework based on the existing deterministic flow solver to examine the influence of uncertain factors on sensitive acoustic phenomena. The focus of this talk will be on our research efforts into non-intrusive methods, which will be accompanied by an approach based on intrusive methods in later stages of the project. We will discuss briefly a variant of the Monte Carlo (MC) method and a spectral projection (NISP) approach. In the MC approach, the stochastic quantities are evaluated on nested levels of spatial discretization, and the combination across these levels is used to achieve a more efficient overall method. In the NISP approach, the quantity of interest is approximated by a polynomial chaos expansion, in which the approximation coefficients are evaluated by (discrete) projection. We will discuss the coupling of both methods to the DG framework and show first applications to a 2D open cavity flow problem in which an acoustic feedback mechanism evolves. In our investigations, we will introduce uncertainties into the geometric configuration (cavity aspect ratio) and flow parameters. At the conference, we will discuss the sensitivity of the hydrodynamics and pressure spectra with respect to these uncertain inputs.

Within the "Digital Human" project we want to contribute to the grand vision of a virtual physiological human model, by developing a realistic model of the neuromuscular system. Due to the high complexity of the neuromuscular system, detailed multiscale computational models are indispensable, e.g. representing the chemo-electro-mechanical behaviour of the system or the global mechanics.

To make progress towards a close-to-realistic model of the neuromuscular system we aim at a holistic revision of the physical, numerical and computational models at all scales. We adapt solution strategies of the discretized models to increase the discretization order and thus, reduce the computational effort. Additionally, we realised parallelization improvements and performed scalability measurements to examine the main bottlenecks.

The work is carried out using OpenCMISS, a code framework for developing computational modelling and visualisation software, particularly targeting bioengineering.

The quality of surface water typically depends upon a complex interplay between physical, chemical and biological factors which are far from being completely understood today. Nevertheless, ensuring an acceptable water quality is a crucial requirement in many environmentally-relevant processes. As a consequence, there still exist large uncertainties related to quality predictions based on state-of-the-art mathematical models of surface water bodies. One of the major shortcomings of the common approaches is their strong simplification with respect to turbulent transport phenomena (often treated with a statistical model and empirically-fitted coefficients) as well as to the multiphase nature of the surface water system. Within the project "microorganisms and turbulence", we are aiming to provide detailed numerical data on the spreading of microorganisms in streaming water bodies, which will later be used to improve simpler models for water quality prediction. Therefore, we conduct several direct numerical simulations (DNS) of turbulent open-channel flow, which are coupled with other physical and microbiological processes. This includes the transport and population dynamics of dissolved and deposited bacteria [1], the transport of nutrients and dissolved gases [2], fully-coupled interactions of finite-size particles with the flow [3], as well as inter-particulate collisions [4]. Since all relevant scales are resolved, we are able to shed light on phenomena which are not readily accessible in simpler models, or even experiments, of classical water science, such as the effect of spatio-temporal heterogeneities on the inactivation of bacteria. In this talk, the modeling methods applied and their computational difficulties will be presented and the role of high-performance computing in such a multidisciplinary will be evaluated. Visualizations of some results of such DNS will be shown and discussed with regard to their implication on the fate of microorganisms. [1] M. T. Auer and S. L. Niehaus, “Modeling fecal coliform bacteria—I. Field and laboratory determination of loss kinetics,” Water Research, vol. 27, no. 4, pp. 693–701, Apr. 1993. [2] B. Kubrak, H. Herlina, F. Greve, and J. G. Wissink, “Low-diffusivity scalar transport using a WENO scheme and dual meshing,” Journal of Computational Physics, vol. 240, pp. 158–173, May 2013. [3] M. Uhlmann, “An immersed boundary method with direct forcing for the simulation of particulate flows,” Journal of Computational Physics, vol. 209, no. 2, pp. 448–476, 2005. [4] A. G. Kidanemariam and M. Uhlmann, “Interface-resolved direct numerical simulation of the erosion of a sediment bed sheared by laminar channel flow,” International Journal of Multiphase Flow, vol. 67, pp. 174–188, 2014.

Our project expands the scope of and merges existing structures for basic HPC programming and tuning support service at Technische Universität Darmstadt, Goethe-Universität Frankfurt, Johannes-Gutenberg-Universität Mainz, and Technische Universität Kaiserslautern. Together, these four sites provide not only a rich pool of challenging HPC applications emerging from first-class science, but also researchers dedicated to the advancement of HPC itself we can capitalize on. Thus, we want to deepen HPC service in areas where scientific expertise at these universities coincides with critical user needs, pushing expertise on GPU performance engineering, model-based scalability analysis, and algorithmic stability and reproducibility into performance engineering practices. The issue of GPU and vectorization performance is crucial, as it is relevant for many HPC architectures. Performance modeling is critical to identify potential performance bottlenecks before resources are expended unnecessarily, an issue that is essential for scaling applications, and, in particular, the efficient transition between tier-3, tier-2, and tier-1 computing centers. Lastly, algorithmic stability and reproducibility are becoming more and more important with respect to the credibility of HPC simulations. In this fashion, we are creating a service organization that offers organizational and operational synergies for HPC users in our two states, exploiting the close spatial proximity of the sites involved. In addition, we are providing a model for merging basic on-site HPC support in a distributed organization with priority support for topics of relevance.

Die Anschaffung und der Betrieb großer HPC Systeme ist kostspielig, daher ist die effiziente Nutzung dieser wichtigen Forschungsinfrastruktur von großer Bedeutung. Diese Aufgabe wird erschwert durch die Komplexität der Systeme und Ihrer Software-Umgebung. Speziell an akademischen Rechenzentren gibt es ein weites Spektrum an Applikationen mit teilweise sehr unterschiedlichen Hardwareanforderungen. Gerade kleine bis mittlere HPC-Zentren haben nur sehr beschränkte Ressourcen für den Support und die Optimierung von Applikationen. Für diese Zentren ist eine effiziente Nutzung der Systeme nur schwer sicherzustellen. Das DFG-Projekt ProPE versucht Komponenten für eine zentrenübergreifende Performance-Engineering (PE) Infrastruktur zu erarbeiten. Dies umfasst einen Prozess zur systematischen Behandlung von PE Fällen und, falls notwendig, deren Delegation innerhalb eines Zusammenschlusses von HPC-Zentren. Des Weiteren beinhaltet es job-spezifisches Performance Monitoring zur Unterstützung der Mitarbeiter um Jobs mit hohem Optimierungspotenzial zu erkennen. Neben einer kurzen Übersicht über das ProPE-Projekt, liegt der Fokus des Vortrags auf Prozess- und Werkzeugaspekten.

While support and training is widely available for users of Tier-1 and Tier-0 computing centers, this is less so for lower-tier centers. The project SES-HPC aims to improve support, both technical and administrative, and training for users of Tier-3 high-performance computing resources, both those who intend to eventually run their applications on higher-tier hardware and those who do not. The project includes the development of a training course program, the support of scientists in developing and optimizing their software as well as running off-the-shelf computing software, administrative help for users who want to switch to higher-tier centers, and the organization of knowledge transfer and documentation. In this talk, the project focus and structure of SES-HPC are introduced.

Monte-Carlo (MC) techniques have a long history with applications in a variety of fields. The most popular approach is the Markov-Chain Monte-Carlo technique where new configurations are proposed based on the current state and proposed step is accepted based on the energy difference. While this technique allows efficient sampling, it often suffers from a very small acceptance rate, driving up the computational cost of this approach. This problem is particularly important in dense systems, such as liquids, glasses and biological systems, where molecular dynamics simulations have become the standard simulation technique. The goal of the MSMEE project in the HPC2 program of BW-Stiftung is to develop a new approach to Monte-Carlo Simulations, aimed at improving the acceptance rate. In this scheme, random steps are weighed based on the local gradient of the potential energy such that beneficial steps will be generated with a higher probability. I will show how this approach can be designed to still fulfill detailed balance while improving the acceptance rate.

Innovations in the automotive industry are driven to a big part by new functionality implemented in software. Many of these innovations are safety-critical. Traditional testing and simulation approaches are becoming less and less capable to achieve the required safety levels at reasonable cost. Thus new approaches are needed. In the HIVES project, funded by the BW-Stiftung, we examine novel approaches that adapt modern quality assurance techniques -- such as static software analysis or virtual prototypes -- for supercomputers. By using new massively-parallel algorithms, even large software systems can be analyzed with highest precision. In this talk we present goals and preliminary results of the HIVES project.

Ziel des DFG-geförderten Projektes ProfiT-HPC ist, das Bewusstsein für Performance-Probleme bei allen Anwendern zu vergrößern. Da sich immer mehr Fachgebiete HPC Ressourcen zu Nutzen machen und diese gleichzeitig immer komplexer werden, machen sich Erfahrungsunterschiede in der effizienten Nutzung der HPC Systeme stark bemerkbar. Um dem entgegen zu wirken, sollen gesammelte Metriken in einem Bericht so aufbereitet werden, dass Nutzer aller Erfahrungsstufen damit einen Ansatzpunkt zum besseren Verständnis und zur Optimierung ihrer Anwendung bekommen. Hierfür wird ein mehrstufiges Modell angestrebt, das in jeder Stufe einen tieferen Einblick in die Anwendung erlaubt, aber hierfür auch ein besseres Verständnis den Nutzers voraussetzt. So können unerfahrene Nutzer Schritt für Schritt an ein Performance Profil herangeführt werden. Der Vortrag gibt einen Überblick über den aktuellen Stand der Arbeiten sowie einen Ausblick auf die bevorstehenden Schritte.

In dem Projekt Mekong soll eine Methodik entwickelt werden, welche eine automatische Skalierung von für einzelne GPUs entworfene Codes auf mehrere GPUs und sogar heterogene Systeme ermöglicht, und dadurch Speicher- und Rechenressourcen aggregiert. Unser Ansatz basiert auf drei Schritten: zuerst werden Informationen wie Datenabhängigkeiten und Speicherzugriffsmuster über statische Code-Analyse gesammelt. Diese Informationen führen über Kostenmodelle zu einer geeigneten Partitionierungsstrategie. Über Code-Transformationen wird dann die gewählte Partitionierungsstrategie umgesetzt. In diesem Vortrag stellen wir kurz die wichtigsten Erkenntnisse und Entscheidungen aus den ersten 10 Monaten des Projektes vor.

In METACCA entwickeln wir neue Ansätze zur Erstellung von domänen-spezifischen Bibliotheken und Sprachen mithilfe von partieller Auswertung. Wir berichten über erste Ergebnisse zweier Anwendungen (High-Performance Ray Tracing, Sequence Alignment aus der Bioinformatik), die wir auf CPU Vektorcode und GPUs abgeildet haben.

Das Social-Event der 7. HPC-Status-Konferenz findet im "Vinum im Literaturhaus" statt.

Anfahrt vom HLRS:
Jeder Bus Richtung Universität, Ausstieg Universität
Jede S-Bahn Richtung Hauptbahnhof, Ausstieg Rotebühlplatz/Stadtmitte
Bus 43 Richtung Killesberg, Ausstieg Berliner Platz oder
U-Bahn U 14, U 2, U 24, Ausstieg Berliner Platz