2008 guest student programme

The 2008 guest student programme ran from 4 August to 10 October 2008 with 10 students.

Group Photo

Persons on the photo, left to right, front to back:
1. row: Ventsislav Petkov, Christoph Honisch, Martin Hoffmann, Walter Nadler, Inge Gutheil, Niklas Fricke
2. row: Tom Schröder, Lutz Roese-Koerner, Dorian Krause, Martin Galgon, Robert Speck
3. row: Matthias Voigt, Stefan Müller, Markus Peschina, Godehard Sutmann, Guido Arnold, Matthias Bolten

Proceedings

Bolten, Matthias (Hrsg.) (2008):
Beiträge zum Wissenschaftlichen Rechnen - Ergebnisse des Gaststudentenprogramms 2008 des John von Neumann-Instituts für Computing (PDF, 5 MB),
Technical Report FZJ-JSC-IB-2008-07

Abstracts

NMR Signal Formation in Capillary Networks

Niklas Fricke, Universität Leipzig
Adviser: Walter Nadler, JSC
Nuclear magnetic resonance (NMR) signals are strongly influenced by local field inhomogeneities. This can be exploited to investigate biological tissue on length scales below the resolution of ordinary magnetic resonance imaging methods and may yield information about size and distribution of capillaries in the myocardium allowing to diagnose stenosis of coronary arteries. However, the actual dependence of the signal on physical parameters (such as capillary width), especially under the influence of diffusion, is only poorly understood, and even for very simplified mathematical models, analytical solutions for the relevant quantities are usually not available. Here, a Monte-Carlo approach was chosen to simulate various models for capillary systems. The frequency autocorrelation functions and the NMR-signals obtained are compared to numerical results for the simplest model in order to test its applicability and the validity of the approximations employed in the numerical calculations.

Visualizing Soil-Plant-Interactions

Martin Galgon, Bergische Universität Wuppertal
Advisers: Tom Schröder, Helmut Schumacher, JSC
With R-SWMS being a complex 3D simulation model for water flow, solute transport and root growth, the need to present the generated amount of data in a clear, easy and understandable way is high. As most simple solutions fail at creating the complex images quickly and without a quality that satisfies scientific needs, a new, more powerful concept is needed. We present RSWMSViz, based on the extensive OpenGL accelerated VTK system. RSWMSViz is a Java program for both, high performance and high quality R-SWMS data presentation, coupled with user interaction and file export possibilities.

Structure Analysis of Communication and Memory Behavior in the Parallel Tree Code PEPC

Martin Hoffmann, Martin-Luther-Universität Halle-Wittenberg
Advisers: Paul Gibbon, Robert Speck, JSC
PEPC - (Pretty Efficient Parallel Coulomb-solver) - is an efficient, portable implementation of a parallel tree code based on the ideas of Barnes and Hut. This code for rapid computation of long-range (1/r) Coulomb forces is presented for use as a 'black-box' library for molecular dynamics applications. Previous experience with the code shows that for some particle configurations some processes have a lot more point-to-point communication than others, ultimately leading to storage imbalance. The causes of this problem are analysed with the help of additional statistical diagnostics and real-time visualisation of the tree structure.

Protein Simulations Under Constraints

Christoph Honisch, Westfälische Wilhelms-Universität Münster
Adviser: Jan Meinke, JSC
Structure prediction of proteins is an important task in biology. There are two principal ways - among others - in which computational methods can contribute to tackle this problem. The first approach consists of computer simulation using molecular dynamics or Monte Carlo based on physical force fields. The second approach uses statistical predictions derived from protein databases of known conformations. In this article a method is proposed to combine the physical and the knowledge based ways of structure prediction by using statistical predictions as constraints in Monte Carlo simulations.

Parallel Scaletransfer in Multiscale Simulations

Dorian Krause, Institute for Numerical Simulation Bonn
Adviser: Bernhard Steffen, JSC
We discuss a recently developed method for coupling molecular dynamics and continuum methods, which is based on a function space approach. The method modifies the bridging domain method by introducing new constraints. We particular focus on the parallel implementation of the method.

Fast Computation of the Ewald Sum

Stefan Müller, University of Technology Chemnitz
Adviser: Godehard Sutmann, JSC
In this report a fast method for computing the Ewald sum for many-particle systems based on the recently developed Fastsum algorithm is proposed. The algorithm is implemented in C and details of the implementation of this algorithm are described. The influence of some parameters on the accuracy of this method is evaluated.

Correcting Erroneous Quantum Algorithms

Markus Peschina, Johannes Gutenberg University Mainz
Advisers: Guido Arnold, Markus Richter, Binh Trieu, JSC
The IQCS (Improved Quantum Computer Simulations) software package is used to simulate Grover's quantum search algorithm. The programming language C, and the machines JUMP and JUGENE are used. The simulation deals with decoherence errors and gate-imperfections and there are four different error stages implemented: ideal, erroneous, ancilla QEC (Quantum Error Correction) and full QEC.

SIONlib - Scalable I/O library for Native Parallel Access to Binary Files

Ventsislav Petkov, Technische Universität München
Adviser: Wolfgang Frings, JSC
High Performance Computing is playing a major role in the research and the development process nowadays. The increasing performance of the computers and the development of the data storage devices open the door for new research fields and discoveries. However, the increasing number of processors used in the supercomputers today poses a new problem to the developers: the scalability of the applications.
One of the major factors for this is the data access and storage. How should the data be written in parallel? Is there an optimized way for doing this? How are the files maintained?
SIONlib is an attempt to answer these questions by providing an easy and efficient interface for working with binary files in parallel.

Preconditioners for the Conjugate Gradient Algorithm

Lutz Roese-Koerner, Universität Bonn
Adviser: Bernhard Steffen, JSC
In this report the Conjugate Gradient Algorithm is explained and the results of the SPARSKIT implementation of this method are shown. Some properties of Incomplete LU Factorization with Threshold are examined and its capability to be used as preconditioner for CG is analyzed. Finally, some attempts are made to detect a problem occurring when using ILUT and CG for certain matrices.

Analysis of the Influence of the Dual Floating Point Unit "Double Hummer" on the Performance of Specific Library Routines on Blue Gene/P

Matthias Voigt, Technische Universität Chemnitz
Adviser: Inge Gutheil, JSC
The Blue Gene/P system of IBM is the newest supercomputer at the Forschungszentrum Jülich GmbH in Germany. It is used for applications in natural sciences that are expensive in terms of computational effort. One of Blue Gene's specialties is the dual floating point unit Double Hummer which is able to perform two floating point operations in parallel. It enables the user to decrease the runtime of applications significantly. During the usage of the dual FPU various problems appeared. The goal of this work is to show what the user should do and what he should avoid to get an optimal performance while using "Double Hummer". We also show how the performance behaves if we modify the code of calling routines or if we use special compiler options. In this way this paper should be a little help for users to get full benefits of the IBM Blue Gene/P architecture.