| CSM-01 |
Approximation of Admissible Measure Valued Solutions for Incompressible Euler Equations Valued Solutions for Incompressible Euler Equations, Filippo Leonardi (ETH Zurich, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-01
Approximation of Admissible Measure Valued Solutions for Incompressible Euler Equations Valued Solutions for Incompressible Euler Equations, Filippo Leonardi (ETH Zurich, Switzerland)
Co-authors: Siddhartha Mishra (ETH Zurich, Switzerland)
We propose a new, first order numerical scheme for the approximation of incompressible Euler equations. This scheme has a number of interesting properties, which mimic the behaviour of the governing equations. We also propose a procedure, which allows a consistent approximation of admissible measure valued solutions. Numerical experiments demonstrate statistical convergence of quantities of interest.
| CSM-02 |
Computing Entries of Inverse Matrices in Genomic Prediction Problems, Fabio Verbosio (Università della Svizzera italiana, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-02
Computing Entries of Inverse Matrices in Genomic Prediction Problems, Fabio Verbosio (Università della Svizzera italiana, Switzerland)
Co-authors: Arne De Coninck (Universiteit Gent, Belgium); Drosos Kourounis (Università della Svizzera italiana, Switzerland); Olaf Schenk (Università della Svizzera italiana, Switzerland)
Genomic prediction problems for plant breeding require scalable and efficient techniques for the solution of sparse linear systems where the matrices involved contain both a large sparse and a dense block. A selective inversion process is suggested that consists of two main steps. First, the selective inverse of the sparse block is computed via the PARDISO solver. Secondly, the Schur-complement of the dense block is computed in parallel and it is used to update the entries of the matrix obtained at the first step. This approach allows the efficient evaluation of selected entries of the inverse, hence its trace that is essential for the solution of the associated maximum likelihood problem.
| CSM-03 |
Discrete Duality Finite Volume (DDFV) Method Applied to COSMO Horizontal Diffusion, Sandie Moody (University of Geneva, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-03
Discrete Duality Finite Volume (DDFV) Method Applied to COSMO Horizontal Diffusion, Sandie Moody (University of Geneva, Switzerland)
Co-authors: Martin Jakob Gander (University of Geneva, Switzerland); Oliver Fuhrer (MeteoSwiss, Switzerland)
At present, the horizontal components of the subgrid scale flux divergence of the averaged equation for mass of water constituents are not being calculated in the COSMO-Model. This is due to the discretization, whos stability is limited. We propose a new approach, namely a coupling of the DDFV and finite volume methods, to be implemented in the COSMO-Model. Our results show that the DDFV method is well adapted to the COSMO-Model as it is stable on any grid type, in particular a terrain-following grid including steep slopes. We also analyse computational costs and convergence rates.
| CSM-04 |
Dynamic Kernel Scheduler (DKS) - a Thin Software Layer Between Host Application and Hardware Accelerators, Uldis Locans (University of Latvia & Paul Scherrer Institute, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-04
Dynamic Kernel Scheduler (DKS) - a Thin Software Layer Between Host Application and Hardware Accelerators, Uldis Locans (University of Latvia & Paul Scherrer Institute, Switzerland)
Co-authors: Andreas Adelmann (Paul Scherrer Institut, Switzerland); Andreas Suter (Paul Scherrer Institut, Switzerland)
Hardware accelerators, such as GPUs and Intel MICs, provide a huge performance potential for HPC applications. However, due to different hardware architectures and development frameworks, taking full advantage of every device and creating manageable code is becoming a challenging task. Dynamic Kernel Scheduler (DKS) provides host application with an interface to schedule communication and task execution on the device and handles all the device and framework specific details necessary to execute these tasks on the accelerator. The concepts and first version of DKS will be presented as well as first results of integrating DKS in different applications (FFT Poisson, MC, Chi-square).
| CSM-05 |
Energy Efficiency of Parareal, Daniel Ruprecht (Università della Svizzera italiana, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-05
Energy Efficiency of Parareal, Daniel Ruprecht (Università della Svizzera italiana, Switzerland)
Co-authors: Andrea Arteaga (ETH Zurich, Switzerland); Rolf Krause (Università della Svizzera italiana, Switzerland)
Parallel-in-time methods have been shown to provide a promising approach to additional concurrency for solving time-dependent problems. However, little is known about their performance in metrics beside speedup. In the future, energy-to-solution for example will become one of the key aspects to judge a method's performance on an HPC system. The poster presents results investigating the Parareal method with respect to its energy efficiency using different parallelization paradigms. A simple theoretical model for the expected energy efficiency is compared against measurements to distinguish between overhead from Parareal's intrinsic limit on parallel efficiency and other sources.
| CSM-06 |
Higher-Order Quasi-Monte Carlo for Bayesian Inversion of Parametric PDEs, Robert Gantner (ETH Zurich, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-06
Higher-Order Quasi-Monte Carlo for Bayesian Inversion of Parametric PDEs, Robert Gantner (ETH Zurich, Switzerland)
Co-authors: Christoph Schwab (ETH Zurich, Switzerland)
We consider multiparametric partial differential equations to which we apply the Bayesian approach to inverse problems. A diffusion equation with distributed, parametric coefficient is considered as model problem. Applications include uncertainty quantification in groundwater flow and shape uncertainty computations. Such problems involve high- or infinite-dimensional integrals, requiring both suitable methods and their massively parallel implementation. We focus here on novel higher-order QMC methods, explaining their efficient construction and their massively parallel application to the problems at hand. Numerical examples consider a parametric space with up to $s=1024$ dimensions.
| CSM-07 |
HPC.m - the MATLAB HPC Compiler and its Use for Solving 3D Poromechanics on Supercomputers, Samuel Omlin (University of Lausanne, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-07
HPC.m - the MATLAB HPC Compiler and its Use for Solving 3D Poromechanics on Supercomputers, Samuel Omlin (University of Lausanne, Switzerland)
Co-authors: Ludovic Räss (University of Lausanne, Switzerland); Yuri Podladchikov (University of Lausanne, Switzerland);
We present a pre-release version of HPC.m - the MATLAB HPC Compiler. HPC.m transforms simple MATLAB scripts in a few seconds into massively parallel near peak performance applications for CPU-, GPU- and MIC-supercomputers. The MATLAB scripts must employ basic syntax and follow a few simple rules. We have successively deployed this software in the development of supercomputing applications for the Piz Daint at CSCS in Lugano, Switzerland. We have shown real-world applications work, as for example a non-linear poro-visco-elasto-plastic 3D solver. Our code achieves near peak performance and scales linearly up to more than 2000 GPUs allowing for a spatial resolution of over 2000^3 grid points.
| CSM-08 |
Time and Energy to Solution Study of the Generalized Eigenvalue Solver, Raffaele Solcà (ETH Zurich, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-08
Time and Energy to Solution Study of the Generalized Eigenvalue Solver, Raffaele Solcà (ETH Zurich, Switzerland)
Co-authors: Thomas Schulthess (ETH Zurich / CSCS, Switzerland)
We study time and energy to solution of electronic structure simulation in terms of a model that relates application performance to machine parameters. We focus on the generalised eigenvalue problems for dense matrices, and study distributed memory architectures with multi-core CPU and hybrid CPU-GPU nodes. The model explains, under certain conditions, the empirically observed affine relationship between node-hours and the energy to solution consumed by the computation. It allows us to extract an effective dynamic energy and static power (which we relate to effective leakage) for the application running on different architectures.
| CSM-09 |
Parallel Solver for the Space Inhomogeneous and Time Dependent Boltzmann Equation, Simon Pintarelli (ETH Zurich, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-09
Parallel Solver for the Space Inhomogeneous and Time Dependent Boltzmann Equation, Simon Pintarelli (ETH Zurich, Switzerland)
Co-authors: Philipp Grohs (ETH Zurich, Switzerland); Ralf Hiptmair (ETH Zurich, Switzerland)
We present a high-performance implementation for the solution of the space
inhomogeneous and time dependent Boltzmann equation. The phase space is discretized using finite elements for the physical domain and a polar spectral discretization based on Laguerre polynomials in velocity. Computations are done in 2+2+1 dimensions with an implicit/explicit split time stepping scheme on unstructured meshes. The polar spectral scheme requires no truncation in velocity and conserves mass, momentum and energy. Dirichlet type boundary conditions are included into a least squares formulation. Results for supersonic gas flows in complicated geometries will be presented.
| CSM-10 |
Pipelined Flexible Krylov Subspace Methods for Large-Scale Computing, Sascha Schnepp (ETH Zurich, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-10
Pipelined Flexible Krylov Subspace Methods for Large-Scale Computing, Sascha Schnepp (ETH Zurich, Switzerland)
Co-authors: Patrick Sanan (Università della Svizzera italiana, Switzerland);
Dave May (ETH Zurich, Switzerland)
We present variants of Conjugate Gradient (CG), Conjugate Residual (CR), and Generalized Minimal Residual (GMRES) methods which are both pipelined and flexible, allowing overlap of global reductions with sparse matrix multiplies and inexact or variable preconditioner applications. The methods are aimed at hiding network latencies for high performance computing using large compute clusters. We demonstrate their effectiveness on synthetic problems and practical examples for nonlinear mantle convection with heterogeneous viscosity structure.
| CSM-11 |
Simulating Large-Scale Scattering Phenomena with the Open-source Boundary Element Library BEM++, Elwin van 't Wout (University College London, United Kingdom) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-11
Simulating Large-Scale Scattering Phenomena with the Open-source Boundary Element Library BEM++, Elwin van 't Wout (University College London, United Kingdom)
Co-authors: Timo Betcke (University College London, United Kingdom); Pierre Gélat (University College London, United Kingdom); Simon Arridge (University College London, United Kingdom)
The scattering of acoustic and electromagnetic waves are important phenomena in many physical models. The boundary element method naturally reduces the wave propagation in exterior domains into a model on the surface of the object. The potential benefits for simulation of large-scale engineering problems can only be achieved with additional fast algorithms. The open-source library BEM++ provides an advanced computing platform for scattering analysis of general 3D structures. We will present ongoing developments and demonstrate its applicability to large-scale scattering phenomena with simulations of high-intensity focused ultrasound modalities for medical treatment of liver cancer.
| CSM-12 |
Snowball Sampling for Modeling Large Networks, Alberto Caimo (Università della Svizzera italiana, Switzerland) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-12
Snowball Sampling for Modeling Large Networks, Alberto Caimo (Università della Svizzera italiana, Switzerland)
Co-authors: Alessandro Lomi (Università della Svizzera italiana, Switzerland); Rolf Krause (Università della Svizzera italiana, Switzerland); Garry Robins (Melbourne School of Psychological Sciences, Australia); Alex Stivala (Melbourne School of Psychological Sciences, Australia); Johan Koskinen (University of Manchester, United Kingdom); David Rolls (Melbourne School of Psychological Sciences, Australia); Peng Wang, Swinburne University of Technology, Noshir Contractor (Northwestern University, USA)
The exponential random graph model (ERGM) is a statistical model for analyzing social networks. However, estimating ERGM parameters is a computationally intensive procedure that imposes severe limits on the size of networks that can be fitted. Recently, it has been shown that conditional estimation can be used to estimate ERGM parameters by estimating parameters for smaller conditionally independent subsets of the network. Snowball sampling can be used to generate such subsets. A large number of relatively small samples can be estimated in parallel, taking advantage of parallel computing to allow estimation of much larger networks than previously possible.
| CSM-13 |
Stencil-Based Exascale Simulations Using an N-Dimensional Array Toolkit, Imen Chakroun (IMEC, Belgium) Abstract |
Poster Session, Monday, June 1, 2015
17:30 - 20:00
CSM-13
Stencil-Based Exascale Simulations Using an N-Dimensional Array Toolkit, Imen Chakroun (IMEC, Belgium)
Co-authors: Imen Chakroun (IMEC, Belgium); Zubair Wadood Bhatti (Katholieke Universiteit Leuven, Belgium); Tom Vander Aa (IMEC, Belgium); Roel Wuyts (Katholieke Universiteit Leuven, Belgium/ IMEC, Belgium); Wolfgang Demeuter (Vrije Universiteit Brussel, Belgium)
The goal of a scientist running simulations on top of exascale systems is to take advantage of the latest HPC technologies which are likely to guarantee better performances. Yet, its goals are to solve scientific problems rather than associated software engineering tasks. We introduce ExaShark, a large-scale n-dimensional array toolkit offered as a high-level library for scientists to help computing exascale simulations. It particularly integrates Patus for preprocessing applications that need stencil computations. While ExaShark handles the distributions of the grid using a plethora of back-end communication protocols, Patus computes optimized stencils on inner regions of the grid.
