2014 guest student programme
The 2014 guest student programme ran from 4 August to 10 October 2014 with 11 students.
Group Photo
Abstracts
Wading the Sea of Isomeres - Topological Aspects and Relative Stability of Allotropomeres of Cage Clusters
Christoph Schran
Adviser: Thomas Müller, JSC
This work presents the automated work flow for an analysis of different isomers of fullerenes and provides results for the connection of topological markers to electronic properties. The work flow was applied to 100 isolated pentagon isomers and non isolated pentagon isomers.
Using Blue Gene compute nodes as accelerator
Alessandro Gabbana
Adviser: Dirk Pleiter, JSC
The constantly growing performance gap between processor, memory and I/O speeds makes it necessary to consider new solutions to enable memory intensive applications exploiting HPC systems. As data-sets continue to grow the traditional storage architectures need to be redesigned. A possible solution consists in moving computational power closer to the data. We will here describe the Blue Gene Active Storage (BGAS), an architecture which aims at tightly coupling computing power and storage. We will discuss how BGAS can be used to perform heterogeneous programming on a large scale using a computational cluster as accelerator. We also describe the implementation of a communication protocol based on active messages in order to establish an interface between BGAS and the cluster. Finally we will address the load balancing problem, an intrinsic factor of both heterogeneous computing and the active messages programming model.
Lambda-Dynamics within the Fast Multipole Method
Andreas Hofmann
Adviser: Ivo Kabadshow, JSC
Molecular dynamics (MD) simulations are essential for the investigation of the properties of biomolecules, such as peptides and proteins. A crucial part of MD simulations is the computation of the long-range electrostatic interactions. With respect to these Coulomb interactions, we currently face two challenges. On the one hand, future exascale computers give rise to the need of highly scalable algorithms. On the other hand, there is a need for more realistic simulations that incorporate dynamic ionisation states (Lambda-dynamics). The Fast Multipole Method (FMM) offers the solution to both problems. This work presents the integration of Lambda-dynamics into two integral parts of the FMM.
SPICOM - Implementing structural plasticity for a full-scale model of the cortical microcircuit
Francesco Rovico
Adviser: Markus Butz-Ostendorf, JSC
Biological evidences point that the brain is not as hard wired as traditionally thought. Neurons are able to locally reorganize their connections to reach an homeostatic value of the electrical activity. Dendrites and axons growth over time is modulated by the need of every neuron to reach an internal equilibrium state.
By adopting the model of structural plasticity, it is possible to grow connectivity in a full-scale cortical microcircuit model with eight different populations, using a constant background noise as input. Such model provides simple rules for synapses formation, deletion and recombination. In conclusion it is of great interest to compare the connectivity that has been grown by the means of structural plasticity and the biological model coming from the literature.
Tools for Preparation and Analysis of Molecular Dynamics Simulations - Applications in Fast Equilibration Procedures
Georgi Stoychev
Adviser: Viorel Chihaia, JSC
When preparing systems for molecular dynamics (MD) simulations, random initial velocities are often used. This approach requires a long equilibration phase before the production run of the simulation. It is possible to assign velocities that are more correlated with the interatomic interactions, which should lead to a faster equilibration. For this purpose a package of tools was prepared, that can be used both to prepare initial configurations and analyse the equilibration of MD simulations. Some of the capabilities of the program are demonstrated for a simple system (diamond).
Parallelization of hp-adaptivity in the FEM library deal.II
Ivan Tonkov
Adviser: Kevin Drzycimski, JSC
The current implementation of the open source finite element library deal.II supports various forms of mesh adaptivity, including "hp". It readily features parallel distribution onto thousands of processors.
However, a combination of the two has not been developed although the question about it has been risen a few times in the past. An attempt has been made to implement that functionality but setting up the working environment and the initial preparations took most of the time available. Therefore, at the present state the "hp"-distributed class is not ready. Future work on the subject will continue and can be monitored on our github repository.
Development and Parallelization of a Generator of Surrogate EEG Data - Implemented in Python
Liudmila Zaytseva
Adviser: Sven Strohmer (JSC)
This projects aims to provide a way to prove the statistical validity of results of the Event-Related Potentials (ERP) registering. Surrogate brain data matched with real EEG generator will allow to more easily identify the presence of ERP-like elements.
ERP are able to provide an opportunity for the effective Brain-Computer-Interface feedback transmission and communication, by informing the effector about the fact that the user's command was interpreted wrongly and thus an error was committed. Statistical significance of ERP elements of artificial data would help to estimate the reliability of ERP appearing in real recordings.
Hybrid Monte Carlo Study of Polymer Chains
Marius Bause
Adviser: Thomas Neuhaus, JSC
The Hybrid Monte Carlo offers a a potentially efficient approach for the simulation of complex systems combining sophisticated Monte Carlo methods with Langevin equation of motion. It is applied to a polymer bead-spring model in three continuous dimensions in Multiple Gaussian Modified Ensembles and its efficiency is compared to Monte Carlo simulation. Phase transitions and quasi-crystalline states of the polymer are analysed.
A Task-Parallel Dual Tree Traversal for the Tree Code PEPC
Marvin-Lucas Henkel
Adviser: Benedikt Steinbusch, JSC
The solution of body problems are a challenging problem of physics because of long ranged potentials and the resulting time complexity of N2. Therefore approximate numerical methods were developed to accomplish an acceptable time complexity. One subset of these are hierarchical methods. The most prominent representatives of this category are the Barnes-Hut tree code and the Fast Multipole Method. In this report the Dual Tree Traversal that makes use of elements of both the Barnes-Hut tree code and the Fast Multipole Method is investigated. The first part of this work contains a description of the algorithm and its parallel implementation into the parallel tree code PEPC. The second part comprises a comparison of the implementation and a Barnes-Hut tree traversal of PEPC.
Towards a hybrid parallelization of Chebyshev Filtered Subspace Iteration - a C++ interface and OpenMP4.0
Marija Kranjcevic
Adviser: Edoardo Di Napoli, JSC
In many scientific applications the solutions of non-linear differential equations are found by solving a number of successive eigenproblems. These eigenproblems are regarded as a sequence whenever the solutions of one problem initialize the next and are said to be correlated if there exists a connection between the solutions of adjacent problems.
This paper shows the Chebyshev Filtered Subspace Iteration method (ChFSI), which exploits that correlation. With a hybrid parallelization in mind, ChFSI is implemented in C++ and parallelized using OpenMP4.0. It is shown that the algorithm performs significantly better than the corresponding direct eigensolver. Offloading is discussed and the code is run on Intel Xeon Phi.
Stochastic and Interaction based Boundary Conditions in MPCD and Flow Simulations
Stephan Schulz
Adviser: René Halver, JSC
In flow simulations of complex geometries on a mesoscopic scale MPCD with traditional boundary conditions is often used amongst others, with the fluid being represented by mesoscopic fluid particles. Boundary artifacts may occour when particles get trapped or even move across boundaries. To circumvent these problems, event-driven boundary conditions are exchanged for interaction contrlled boundary conditions. Interaction with the boundary takes the form of a Lennard-Jones potential, stochastic or modified bounce-back rules within an interaction zone.
These new boundary conditions create a density excess in front of the interaction zone or for the potential based boundary conditions within the potential. This density excess is influenced by the form of the potential and cannot be eliminated unless the treatment of fluid particles close to the boundary inside the interaction region is modified.
Poster announcing the 2014 Guest Student Colloquium