# Guest Student Programme 2017

## Abstracts

### Automatic Classification of Neural Network States - Machine Learning for Simulated Neuroscientific Data

**Avleen Sahni, Physikalisches Institut, Rheinische Friedrich-Wilhelms-Universität Bonn, Germany**

**Adviser: Wouter Klijn**

The visual cortex is a central part of the cerebral cortex of the brain, responsible for processing visual signals. At a high level of abstraction, the cortex can be seen of a layered network of neurons with a 2D topology. It has been observed that neurons with such a topology display multiple self-organising dynamic states, when they fire without external input. In this project, these multi-stable dynamic states were identified and classified automatically using machine learning algorithms. The data was in the form of images containing binned and time averaged neuron spikes of 20 million neurons, which were preprocessed into a form that was amenable for machine learning. The performance of different classification algorithms was compared and it was found that for the given data set, the k-nearest neighbour algorithm worked best.

Reducing Cache Coherence in a Task-Based FMM - MCS Locks in Multithreaded Applications

**Carl-Martin Pfeiler, Institute for Analysis and Scientific Computing TU Wien, Austria**

**Adviser: David Haensel**

Synchronization barriers are essential for the correctness of multithreaded scientific applications. Utilizing shared flags for synchronizing threads is the standard approach. We consider a task-based modern C++ implementation of the Fast Multipole Method (FMM) and argue that employing locking strategies relying on thread local flags for intra-node synchronization clearly yields a reduction of cache coherence. Our observations are underpinned by benchmarks which show a significant improvement of the application’s performance.

Finite Element Solution for the Acoustic Scattering Problem - Adaptive Mesh Refinement using deal.II

**Duygu Kan, Computational Science and Engineering Program, Istanbul Technical University, Turkey**

**Adviser: Marc Fehling**

In this work acoustic scattering problem related to radially inhomogeneous spheres is investigated. Acoustic scattering problem is solved in 2-D and 3-D by using Finite Element Method (FEM) with Adaptive Mesh Refinement (AMR) using deal.II on High Performance Computing (HPC). Simulations show that the method gives satisfactory resolution and also provides benefit in terms of computation time due to the mesh refinement method used. Moreover appropriate choice of the functions that determine the acoustic profile of the spherical object are of significant importance in the simulations.

Machine learning for iterative system-solves arising within FEAST

**Gayatri Čaklović, Department of Mathematics, University of Zagreb, Croatia**

**Adviser: Edoardo Di Napoli**

Convergence of iterative methods GMRES and MINRES has been a long standing subject of reserach since the end of the 20th century. The aim of this paper is to investigate the possibility to predict the convergence for GMRES and MINRES for non preconditioned Hermitian linear systems using machine learning regression. Different approaches have been tried out and discussed. Results show that a clustering approach seems to be almost necessary. A more in-depth mathematical theory

was presented which clarifies on what does the convergence depend on. At last, this motivates a possible cheaply computable solution to the problem.

Finite volume effects of electrodynamics on the lattice

**Gaspare Di Fede, Department of Physics, University of Pisa, (PI), Italy**

**Adviser: Kalman Szabo**

Here we report on our investigations of finite volume effects in electrodynamics. We first derive the classical result. Then the implementation of quantum electrodynamics on the lattice is reviewed, and details of numerical simulations are given. We finally compare the classical and quantum results.

Deep Learning in Hyperspectral Remote Sensing

**Julius Lange, Department of Physics, Humboldt University of Berlin, Germany**

**Adviser: Morris Riedel**

Convolutional neural network deep-learning techniques are applied to the hyperspectral land cover type classification problem, using the Indian Pines dataset. Unlike the common approach which considers only a subset of the data, we apply our method to the complete set covering all labelled pixels and classes. While many previous works required some form of complex feature engineering to obtain good results, our method is based on an autonomous feature learner. Our results improve on the state of the art accuracy.

HMC simulations of Carbon Nanotubes

**Johann Ostmeyer, Helmholtz-Institut für Strahlen- und Kernphysik, Rheinische Friedrich-Wilhelms-Universität Bonn, Germany**

**Adviser: Stefan Krieg**

This work presents three methods to speed up Hybrid Monte Carlo simulations of carbon nanotubes. First several symplectic integrators are presented and compared. The Omelyan integrator with tuning parameter ζ = 0.193 is found to be the fastest. Next, the initially used CG solver is replaced by a flexible GMRES solver with single precision CG preconditioning. Last, but not least, Hasenbusch acceleration is adapted to the problem introducing one or more staggered mass terms and corresponding additional time scales. This leads to a speed up of approximately factor 13 for large systems. It is believed that this number can easily be increased further if the runtime dependence on many staggered masses is investigated in more detail.

Computing Dirichlet eigenvalues using the MFS - A mesh- and integration-free approach

**Janosch Preuß, Institute for Numerical and Applied Mathematics, University of Göttingen, Germany**

**Adviser: Andreas Kleefeld**

Calculating Dirichlet eigenvalues with high accuracy is a challenging task in scientific computing. For this purpose we implement an algorithm which is based on a modified method of fundamental solutions. Numerical experiments demonstrate that our implementation produces very accurate results for simple domains but may need improvement when the domain becomes more complicated especially in three dimensions.

Lattice QCD Optimization for KNL Architecture - Without Intrinsics!

**Loris Gallina, Dipartimento di Fisica “E. Fermi”, Università di Pisa, Italy**

**Adviser: Eric Gregory**

Lattice QCD is one of the principal approaches to compute physical quantities related to strong interactions. It has been used, for example, to test the agreement between QCD theory and experimental data. However, it is computationally expensive. Therefore it is important to develop methods which can improve the performances of LQCD calculations. In this report some efforts to develop optimization of the calculation of the D operator (part of the fermion matrix) with Intel KNL Xeon Phi Architecture are explained. Vectorization with SIMD (512 bits long) registers and without intrinsic operations, the site-fusion method and the 2-column-compression algorithm are used, which have allowed us to achieve high values of the performance (almost 400 GFlops/s). Finally some ideas for further potential optimization are shown.

Parallel-in-time integration using XBRAID - A study on coupling XBraid and DUNE

**Paras Kumar, Department Informatik, University of Erlangen-Nürnberg, Germany**

**Adviser: Robert Speck**

Employing the recently developed parallel-in-time integration methods in addition to the already well established space parallel solvers has come up as an exciting option for achieving faster compute speeds during the solution of time dependent problems described by partial differential equations. This work aims at developing a time parallel solver for the heat equation by coupling the XBraid algorithm for parallelization in time along with a finite element discretization based on the DUNE framework.

###

Domain Decomposition with Space-filling Geometries

**Saskia Körning, Faculty of Chemistry and Biochemistry, Ruhr-Universität Bochum, Germany**

**Adviser: Godehard Sutmann**

Domain decomposition is a popular method to speed up molecular dynamics simulations. The communication between nodes, due to transfer of particle information, takes time and induces an overhead. This project examines the influence of the domain geometry on communication time. Different shapes for domain geometries were tested: cube, rhombic dodecahedron and truncated octahedron. The rhombic dodecahedron and truncated octahedron were found to have faster communication times than the cube.

### Porting and Effectively Profiling High-Performance Software Experience with tmLQCD on JURON and JULIA

**Yishai Oltchik, The Rachel and Selim Benin School of Computer Science and Engineering, The Hebrew University of Jerusalem, Israel**

**Adviser: Dirk Pleiter**

We take an existing codebase used for Quantum Chromodynamics simulations and install it on JURON and JULIA, two new supercomputer systems with highly different architectures. After porting the codebase, we introduce and demonstrate the use of sampling-based profiling tools in an exploratory attempt to detect performance bottlenecks in the code; in particular focusing on code segments which executed purely on the CPU. Such bottlenecks, once identified, present useful starting points for any further optimization and fine-tuning efforts.

## Group Photo

Copyright: Forschungszentrum Jülich