Catálogo de publicaciones - libros

We discuss performance characteristics of scientific applications on modern computer architectures, ranging from commodity “off-the-shelf” (COTS) systems like clusters, to tailored High Performance Computing (HPC) systems, e.g. NEC SX6 or CRAY X1. The application programs are selected from important HPC projects which have been supported by the KONWIHR project cxHPC. In general we focus on the single processor performance and give some optimisation/parallelisation hints, if appropriate. For computational fluid dynamics (CFD) applications we also discuss parallel performance to compare COTS with tailored HPC systems. We find, that an HPC environment with a few tailored “central” high-end systems and “local” mid-size COTS systems supports our users' requirements best.

In the Peridot project our goal is a portable environment for performance analysis for terascale computing that realizes a combination of new concepts including distribution, on-line processing and automation. In this paper we present the lightweight dynamic application monitoring approach that forms the basis for this environment. In our distributed monitoring solution we try to minimize the perturbation of the target application while retaining flexibility with respect to configurability and close-to-source filtering and pre-processing of performance data. We achieve this goal by separating the monitor in a passive monitoring library linked to the application and an active component called runtime information producer (RIP) which provides performance data (metric and event based) for individual nodes of the system through a monitoring request interface (MRI). By querying a directory service, tools discover which RIPs provide the data they need.

The KONWIHR project has developed a framework for the integration of simulation and visualization for large scale applications. This framework provides its own grid structure, the so called hierarchical hybrid grid, which is well suited for runtime efficient realization of multilevel algorithms. Furthermore, it offers flexible visualization functionality for both local and remote use on number crunchers and workstations. It is based on modern object-oriented software engineering techniques without compromising on performance issues.

In deformations of polynomial functions one may encounter “singularity exchange at infinity” when singular points disappear from the space and produce “virtual” singularities which have an influence on the topology of the limit polynomial. We find several rules of this exchange phenomenon, in which the total quantity of singularity turns out to be not conserved in general.

In this contribution, fully three-dimensional models are used for the numerical simulation of both the structure and the fluid in fluid-structure interaction computations. A partitioned, but fully implicit coupling algorithm is employed. As an example, the wind-excitation of a thin-walled tower is investigated.

We discuss single particle dynamics of the half-filled 2 Hubbard model at → 0 calculated within the dynamical cluster approximation, using numerical renormalization group as non-perturbative cluster solver, which requires the use of parallel architectures with large number of processors and memory. In addition, fast temporal storage for large out-of-core matrices is needed. The results obtained indicate that the half-filled 2 Hubbard model at → 0 is a paramagnetic insulator for values of the Coulomb interaction in strong contrast to weak-coupling theories.

We study the scaling properties of the quantum projected (5) model in three dimensions by means of a highly accurate Quantum-Monte-Carlo analysis. Within the parameter regime studied (temperature and system size), we show that the scaling behavior is consistent with a (5)-symmetric critical behavior in the numerically accessible region. This holds both when the symmetry breaking is caused by quantum fluctuations only as well as when also the static (mean-field) symmetry is moderately broken. We argue that possible departure away from the (5) - symmetric scaling occurs only in an extremely narrow parameter regime, which is inaccessible both experimentally and numerically.

The paper is concerned with the prediction and analysis of the turbulent flow past an unswept NACA-4415 airfoil at high angle of attack. The predictions were carried out using large-eddy simulations (LES) applying two different subgrid-scale (SGS) models, namely the Smagorinsky model and the dynamic model by Germano/Lilly. For this kind of flow simulations high-performance computers such as the presently used SMP cluster Hitachi SR8000-F1 are inevitable. The Reynolds number investigated is = 10 based on the chord length of the airfoil. An inclination angle of = 18° was chosen. At these operating conditions, the flow past the airfoil exhibits a trailing-edge separation including some interesting flow phenomena such as a thin separation bubble, transition, separation of the turbulent boundary layer and large-scale vortical structures in the wake. Qualitatively the simulations with both SGS models predict the aforementioned flow features in a similar manner. However, looked at closely, some noteworthy differences become evident. The most striking one concerns the shape and influence of the separation bubble. In the simulation with the Smagorinsky model the separation bubble is predicted more than twice as thick as by the dynamic model. This also influences quantitative values such as the distribution of , or the turbulent kinetic energy. The largest discrepancies between the results of the two models applied are found to be close to the wall. Therefore, the SGS models have to be examined with respect to their reliability in predicting the near-wall region of a flow. In addition, the paper aims at a deeper insight into the nature of turbulent separated flows. This is done by analyzing the simulations according to the anisotropy-invariant theory which is expected to provide an improved illustration of what happens in a turbulent flow. Therefore, the anisotropy of various portions of the flow was extracted and displayed in the invariant map in order to analyze the state of turbulence in distinct regions. Thus, turbulence itself as well as the way it is developing can be investigated in more detail leading to an improved understanding of the physical mechanisms.

Direct numerical simulations (DNS) of compressible supersonic channel flow of air at Reynolds numbers ranging from = 180 to = 560 and Mach numbers ranging from = 0.3 to = 3.0 have been performed. A Navier-Stokes solver of high order accuracy has been vectorized and parallelized to run efficiently on the Hitachi SR8000-F1. Budgets of the Reynolds stresses and the passive scalar fluxes are presented, as well as explanations concerning the reduction of the pressure-correlation terms, using a Green's function approach.

A Direct Numerical Simulation (DNS) of turbulent channel flow of dilute suspensions of small, Brownian fibres in a Newtonian solvent is presented. The DNS investigates the potential of drag reduction under situations, where no internal elasticity of the additives is present. The DNS is solving the microscopic equations for the suspended fibres and couples the resulting stresses into a (macroscopic) DNS of the solvent. The microscopic equations for the conformation of the fibres as well as the resulting stresses are derived by the rheological theory of dilute suspensions of Brownian particles in Newtonian solvents. These equations are solved by a Monte-Carlo method. First results show a dramatic reduction of the Reynolds shear stress. However, only a mild reduction of the drag is observed because the fibres generate considerable shear stress components at the wall at the configuration chosen.