Search

link to homepage

Institute for Advanced Simulation (IAS)

Navigation and service


Guest Student Programme 2006

Proceedings

Esser, Rüdiger (Hrsg.) (2006)
 Beiträge zum Wissenschaftlichen Rechnen - Ergebnisse des Gaststudentenprogramms 2006 des John von Neumann-Instituts für Computing (PDF, 5 MB)
Technical Report IB-2006-14, 2006

Abstracts

Advanced Memory Access Scheme for Quantum Computer Simulation

Mathias Aust, Universität Leipzig
Adviser: Guido Arnold, Marcus Richter

Starting from an already parallelized simulation code for an ideal quantum computer, we implement a more flexible communication scheme. We give a brief introduction into the necessary concepts and notations for quantum computers and discuss the problems of quantum computer simulations. The fundamental concept for parallelization of the code is explained. Based on this concept benefits and problems of the old and new communication schemes are discussed. Finally, both schemes are compared in different benchmarks.

Pedestrian Dynamics with Event-Driven Simulation

Mohcine Chraibi, Technical University Hamburg-Harburg
Adviser: Armin Seyfried

For the modelling of pedestrian dynamics we treat persons as self-driven objects moving in a continuous space. On the basis of a modified social force model we qualitatively analyze the influence of various approaches for the driving force of the pedestrians on the resulting velocity-density relation. With an event-driven algorithm we will obtain the following result: if the model increases the required space of a person with increasing and self-adapting current velocity, the reproduction of the typical form of the fundamental diagram is possible.

Implementation of a Tabulated Barnes-Hut-Ewald Algorithm for Periodic Boundaries in PEPC

Falk Eilenberger, Universität Jena
Adviser: Paul Gibbon

The parallel Barnes-Hut program PEPC was augmented to support periodic boundary conditions. In consequence PEPC now supports a tabulated Barnes-Hut-Ewald mode, allowing periodic boundary simulations for up to and beyond 106 particles. An analysis of errors with respect to different parameter sets was carried out. It could be shown that the additional error introduced by the periodic boundary conditions can be made small or comparable to the intrinsic Barnes-Hut error with reasonable effort. A new interpolation scheme has been developed that achieves higher accuracy and is key for the success of low error simulations in the range of ϑ ≤ 0.3. It could also be shown that the code has to do less communication per time, which will improve its scalability to higher numbers of CPUs. A strategy for reducing expensive load operations arising from the Ewald tabulation is also proposed.

Ein Beitrag zur Entwicklung einer parallelen Version von AOFORCE des Programmpakets TURBOMOLE

Sebastian Höfener, Universität Karlsruhe
Adviser: Thomas Müller

Für die Chemie sind Strukturen chemischer Verbindungen und deren Reaktivität von grundlegender Bedeutung. Die Berechnung der molekularen Hesse-Matrix bezüglich Kernverrückungen stellt eine Herangehensweise für beide Problemstellungen dar. Diese erlaubt die Berechnung von Schwingungsspektren sowie die Suche nach übergangszuständen, die während chemischer Reaktionen auftreten. Eine effiziente Implementierung befindet sich in dem sequentiellen Modul AOFORCE des Quantenchemie-Programmpakets TURBOMOLE, welches als Ausgangspunkt verwendet wird. Nach Analyse der zeitintensiven Programmteile wurde eine Parallelisierung mittels Global Arrays vorgenommen. Der grundlegende Algorithmus wurde dabei nicht verändert. In verschiedenen Programmteilen wurde dynamisches und statisches Loadbalancing implementiert.

Load Balance and Scalability of Force Decomposition Methods in Parallel Molecular Dynamics Simulations

Florian Janoschek, Universität Stuttgart
Adviser: Godehard Sutmann

Force decomposition methods are a powerful way of parallelization of molecular dynamics simulations. I implemented three different methods introduced by Murty et al. and Taylor et al. For the so called "force-stripped row" method, a load balancing mechanism capable to deal with inhomogeneous systems and an optimized force communication scheme was developed. Both load balance and efficient communication are preconditions for reasonable parallel speedup. All methods were tested on a homogeneous and an inhomogeneous benchmark system in respect of load balance and speedup.

Latency Optimized Parallelization of Near-Field Interactions in the Fast Multipole Method

Bernhard Kühnel, Chemnitz University of Technology
Adviser: Holger Dachsel, Ivo Kabadshow

This report presents an approach to implement the Fast Multipole Method (FMM) on parallel computers. After introducing the FMM, a fast summation algorithm that reduces the complexity of the Coulomb problem from O(N2) to O(N), an implementation of the algorithm and parallelization by dynamic load balancing will be discussed. Finally, a new approach using static load balancing and communication overlapping is introduced. This new scheme is based on Global Arrays and intends to reduce the overall number of communication steps to a minimum, since the restricting part of the parallelization is introduced by the latency.

Determination of Ground State Degeneracy of Systems with Phase Transitions of First Order and a Simple Chain Model Using Improved Parallel Monte Carlo Methods

Martin P. Magiera, Universität Duisburg - Essen
Adviser: Ulrich Hansmann, Thomas Neuhaus

The determination of ground states of systems with many degrees of freedom and their degeneracy is not only an interesting topic in statistical physics, but also in computational biophysics (protein folding) or financial market research. For such systems energy landscape becomes rough and the minimization of their Hamiltonian nontrivial, so statistical methods become necessary. In this work simple systems with phase transition of second and first order (Ising and Potts model) are studied by Monte Carlo simulations, using parallel tempering and multiple Gaussian Monte Carlo methods to make a random walk through the complete phase space possible. The trivial degeneracy of these systems is calculated, simulations are performed parallel on "Jülich Multiprocessor" (JUMP). Finally simple two dimensional chain systems are simulated and their ground state degeneracy is determined, where a new phase transition like behavior is found.

Performance-Messungen eines parallelen Jacobi-Davidson Eigenwertlösers

Frank Schmidt, Technische Universität Chemnitz
Adviser: Bernhard Steffen, Oliver Bücker

Das Jacobi-Davidson Verfahren wurde 1996 vorgestellt. Seither beschäftigt sich die Fachwelt mit Aussagen zur Stabilität und Effizienz des Verfahrens. In dieser Ausarbeitung wurden einige Untersuchungen zu verschiedenen Parametereinstellungen für einen bereits vorhandenen parallelen Jacobi-Davidson Löser durchgeführt und dessen Stabilität getestet.

Parallele Performance der Matrix-Vektor-Multiplikation unter Einsatz verschiedener Speicherformate

Michèle Wandelt, Bergische Universität Wuppertal
Adviser: Bernhard Steffen

In diesem Artikel wird die parallele Performance der Matrix-Vektor-Multiplikation für dünnbesetzte Matrizen in verschiedene Speicherformaten untersucht. Dabei wurde so vorgegangen, dass die Matrix-Vektor-Multiplikation einem im ZAM entwickelten Programmpaket entnommen, angepasst und erweitert wurde. Dieses Programmpaket wurde im Rahmen einer Dissertation in FORTRAN implementiert und im Verlaufe einer Diplomarbeit erweitert.

Rigide und nicht-rigide Bildregistrierung mit dem Insight Toolkit

Markus Weigel, Universität zu Lübeck
Adviser: Uwe Pietrzyk, IME

Diese Ausarbeitung gibt eine Einführung in die rigide und nicht-rigide Bildregistrierung für medizinische Anwendungen und zeigt, wie diese im Insight Segmentation and Registration Toolkit (ITK) umgesetzt sind. Als spezielles nicht-rigides Verfahren wird der Algorithmus von Thirion vorgestellt und seine Umsetzung in ITK untersucht. Das Fernziel eines verallgemeinerten Registrierungsframeworks wird diskutiert.

Group Photo

Guest Students 2006Guest students 2006 and some of their advisers

Persons on the photo, left to right, front to back:

1. row: Rüdiger Esser (ZAM), Mathias Aust (Leipzig), Michèle Wandelt (Wuppertal), Frank Schmidt (Chemnitz), Armin Seyfried (ZAM), Mohcine Chraibi (Hamburg)
2. row: Godehard Sutmann (ZAM), Bernhard Kühnel (Chemnitz), Falk Eilenberger (Jena), Martin P. Magiera (Duisburg), Thomas Müller (ZAM), Bernhard Steffen (ZAM), Florian Janoschek (Stuttgart)
3. row: Markus Weigel (Lübeck), Ivo Kabadshow (ZAM), Guido Arnold (ZAM), Thomas Neuhaus (NIC)


Servicemeu

Homepage