Search

link to homepage

Institute for Advanced Simulation (IAS)

Navigation and service


Guest Student Programme 2015

Abstracts

Hybrid model for pedestrian simulations - Approaches for modeling the transition area

Andrijana Brkić, University of Zagreb
Adviser: Gregor Lämmel

Pedestrian simulations exist in different spatiotemporal precision. The more detailed the model, the complexity of computation and preparation increases. To get better computational efficiency, but in the same time keep the accuracy of the model, a hybrid model is proposed - a detailed (microscopic) model should be applied in areas of greater interest while a coarser (mesoscopic) model is used for the surroundings and areas where not many interactions are expected. To reduce the effect of coupling two models, a transition area is set in between. The pedestrians inside the transition area exist in both models at the same time, which enables the communication between the models. This work introduces an approach for positioning incoming pedestrians in the transition area using the Voronoi diagram and compares it to the naive grid approach.


Local electron Correlation Techniques for Multireference Methods

Christian Schwermann, WWU Münster
Adviser: Dr. Thomas Müller

Proper electron correlation treatment is crucial for an accurate description of chemical and physical properties of molecules, but are mostly limited to smaller systems due to the steep power-law scaling they exhibit. In this work, a local technique based on a simple geometric description of localized molecular orbitals is used to drastically reduce the number of configuration state functions in MRCISD calculations and achieve linear scaling with respect to the size of the -n--particle basis. A series of calculations on different systems confirms the expected scaling, while the error margins are kept within chemical accuracy. Additional calculations exploring a screening procedure based on Coulomb repulsion integrals indicate less balanced results.


Performance Analysis of PFASST

Fabian Köhler, TU Dortmund
Adviser: Torbjörn Klatt

The PFASST introduced in 2012 is an iterative method that allows the numerical solution of ODEs and PDEs with temporal parallelization. When coupled to spatial parallelization strategies, this permits efficient scaling to much higher numbers of processing units. Thereby it is possible to exploit the resources of upcoming HPC systems more effectively. In this report, a general introduction into PFASST is given. Afterwards it is shown how the performance of the PFASST implementation was improved by applying both serial as well as parallel profiling tools.


A Phinal word on tuning of LQCD

Giorgio Silvi, University of Ferrara
Adviser: Stefan Krieg

Lattice QCD is a well-established quantitative approach that has succeeded in calculating many physical quantities, making QCD the only candidate for the theory of the strong interactions. The computational requirements of LQCD demand the usage of supercomputers, as well as methods to improve performance. The Intel Xeon Phi and SIMD vectorization represent a good candidate to fullfill this request. The purpose of this report is to investigate the usage of them in the context of Lattice QCD, with a particular focus on optimization for data management.


Methods for Fast Thermal Equilibration in Molecular Dynamics Simulations

Julian Keupp, Ruhr-Universität Bochum
Adviser: Viorel Chihaia

Molecular dynamics simulations require large amounts of computational resources during the equilibration phase. In order to minimize the computing time for the equilibration, some methods of assigning initial velocities are proposed and analyzed. The code Tools4Lammps was developed as a tool for equilibration characterization and for generating the initial velocities in different ways. The thermal equilibration techniques were tested by NVE MD simulations on a 3x3x3 supercell of methane clathrate. The analysis of different parameters suggest that the standard approach can be improved.


Adhesive friction in sliding rough, elastic bodies

Kaloyan Genkov, Sofia University
Adviser: Martin Müser and Wolf Dapp

The mechanical contact between two solids with rough surfaces occurs on a small fraction of their aparrent contact area. Due adhesion and elastic deformation, the total energy of creating and breaking of a single contact may be negative which can lead to friction. In this work the Green's function molecular dynamics method is used to simulate push-pull motion of two solids with self-affine surfaces. A hysteresis of the mean displacement in the adhesive case is shown, which implies that the system is likely to exhibit adhesive friction.


Coupling continuum elasticity with atomistic fluid dynamics

Lukas Engelke, University of Duisburg-Essen
Adviser: Martin Müser and Wolf Dapp

In order to simulate fluids in the tight gap between rough, elastic bodies, we couple Molecular Dynamics (MD) and Green's Function Molecular Dynamics (GFMD). The fluid is described as particles, which interact by a Lennard-Jones potential, whereas the surfaces of the solids represent exponentially repulsive walls. This enables us to investigate interactions of the fluid and the walls and furthermore to study effects beyond Newtonian fluids. As a first step this method is applied to a problem similar to the flow around a cylinder revealing qualitatively similar results.


Combining Linked-Cell Lists and Verlet Lists in OpenMP

Malte Jongmanns, Universität Duisburg-Essen
Adviser: Godehard Sutmann and Rene Halver

A parallel implementation of a neighbour list technique for Molecular Dynamics is presented. The technique reduces the number of distance calculations between particles and combines the advantages of linked-cell lists and Verlet lists. The parallel implementation approach is described in detail and performance results such as benchmarks and scaling behaviour are discussed. Furthermore, a parameter study of the Verlet list's skin radius is conducted to find the minimum of computational time. It turns out that the presented algorithms show good parallel efficiency.


Latency-avoiding communication schemes for the FMM

Martin Ueding, Rheinische Friedrich-Wilhelms-Universität Bonn
Adviser: David Haensel and Andreas Beckmann

The Fast Multipole Method (FMM) is an O(N) Coulomb force computation scheme suitable for Molecular Dynamics (MD) simulations. It is based around separation of near and far field and uses a multipole expansion for the latter. A serial version of the first steps in the algorithm was parallelized and then analyzed for its scaling behavior.


NEST I/O - Strategies for a peta-scale neural network simulator

Sebastian Billaudelle, Heidelberg University
Adviser: Alexander Peyser

Several efforts like the Human Brain Project work towards brain-scale neural network simulations. With increasing model size and parallelism, the storage of recorded neurophysiological data represents a challenge to the respective I/O infrastructure. Traditional strategies such as task-local ASCII files fail at the filesystem's metadata wall and lead to extreme delays on initialization. This report introduces a novel and flexible approach to I/O for the spiking neural network simulator NEST. As an addition to the legacy approach, we will present a scalable backend based on SIONlib. Both implementations are compared in custom benchmarks as well as first simulations of a cortical microcircuit model.


BGAS as semi-persistent cache

Sara Schiesaro, University of Ferrara
Adviser: Dirk Pleiter

Motivated by the bursty behavior usually observed for I/O on HPC systems, Burst Buffers (BBs) seem to be a promising approach to the I/O challenge in nowadays and future HPC systems. This article discuss a possible use case for Blue Gene Active Storage (BGAS) as a semi-persistent cache to speed-up predominantly reading applications. First, an overview of algorithms used for the implementation of that concept is provided, following with a more detailed description of the ORACLE Berkeley DB tool, how it has been configured and the related implementation challenges/open problems.

 

Adaptive numerical integration in functional Renormalization Group methods

Toni Vidović, University of Zagreb
Advisor: Edoardo Di Napoli

We propose a numerical integration method for computing the right hand side in flow equations for many-fermion systems. We introduce the adaptive integration and make a case for parallelization. We furthermore present results of our algorithms, and suggest future applications.

Group Photo

Guest Student Programm 2015 Photo

left to right:
Christian Schwermann, Julian Keupp, Giorgio Silvi, Sara Schiesaro, Toni Vidović,
Sebastian Billaudelle, Andrijana Brkić, Malte Jongmanns, Lukas Engelke,
Kaloyan Vladimirov Genkov, Martin Ueding, Fabian Köhler


Servicemeu

Homepage