Donnerstag, 25. Oktober 2012 (WIAS, ESH) |
10.50 Uhr |
Eröffnung (Dr. Jürgen Fuhrmann)
|
11.00 Uhr - 11.45 Uhr |
Felix Anker (WIAS Berlin)
Erste Erfahrungen in der GPU-Programmierung
Abstract:
Der Vortrag führt in die Konzepte der GPU-Programmierung ein.
Es wird kurz auf den Aufbau aktueller Grafikkarten eingegangen, welcher
Auswirkungen auf die Programmierung und Performance hat.
Anhand von Code-Beispielen aus den CUDA-Examples wird die grundlegende
Struktur eines Programms erarbeitet und mögliche Flaschenhälse
aufgezeigt. Als Anwendungsbeispiel dient das R-Paket aws. An ihm werden
verschiedene Implementierungen durch OpenMP, CUDA und OpenCL in ihrer
Handhabung und Performance gegenübergestellt.
|
11.45 Uhr - 12.30 Uhr
|
Dr. Carsten Brée (WIAS Berlin)
An MPI-parallelized pseudospectral splitstep method for numerical simulation
of optical filaments
Abstract:
An MPI-parallelized numerical scheme for the propagation of intense optical pulses
in laser filaments is discussed. Mathematically, the evolution of optical pulses in filaments is described
by a unidirectional approximation of Maxwell's equations. This approximation yields a first-order PDE
in the coordinate longitudinal with respect to the beam axis. Numerically, it can be solved
and implemented as a pseudospectral split-step scheme very similar to the numerical method typically
used for the Nonlinear Schrödinger Equation known from fiber optics. However, in contrast
to the fiber optical problem, the problem discussed here is essentially a 2+1-dimensional one.
As MPI follows a distributed-memory approach, this requires efficient means of distributing
and transposing a 2d array of numerical data among the CPU cores involved.
|
12.30 Uhr - 14.00 Uhr
|
Pause |
14.00 Uhr - 14.45 Uhr
|
Dr. Klaus Gärtner (WIAS Berlin)
Die schöne Illusion vom parallelen Rechnen als Lösung aller Probleme
Abstract:
Es wird ein simples Computer-Modell angenommen und versucht,
einfachste Konsequenzen für Aufgaben und extraordinäre Aufgaben
abzuleiten.
Dabei werden folgende Punkte aus der Sicht numerischen Rechnens berührt:
Daten, Rechenwerk, Speicher, Konflikte bei parallelen Aufgaben,
Synchronisation,
Load-Balance, Programmiermodelle und Dimension des Fehlerraumes,
der geometrische Missfit der Gegenwart.
Techniken für die Parallelisierung von Algorithmen ausreichend hoher
Komplexität.
|
14.45 Uhr - 15.30 Uhr
|
Dr. Jürgen Fuhrmann (WIAS Berlin)
OpenMP und parallele Strategien für PDE-Lösungen auf unstrukturierten
Netzen
Abstract:
Der Vortrag diskutiert mögliche Varianten paralleler Lösungsstrategien
für partielle Differentialgleichungen auf unstrukturierten Netzen.
Die Parallelisierung für Shared-Memory-Systeme auf der Basis von
OpenMP wird als Optimum in Bezug auf Programmieraufwand,
Portabilität und erzielbarer Effizienz angesehen.
Des weiteren werden grundlegende Elemente des Konzepts von OpenMP
beschrieben sowie in pdelib realisierte parallele Verfahren
vorgestellt.
|
R. Richter (MPI Halbleiterlabor, München)
Design and simulation of a silicon radiation detector for the Belle II
experiment
Semiconductor radiation detectors exploit the photoelectric effect for
light or particle detection in which the detector performance is given
by the signal to noise (S/N). The obtainable signal is given by the
energy necessary to generate one electron-hole pair which is in silicon
only 3.6eV. With modern silicon technologies very low noise detectors
can be fabricated.
The Depleted Field Effect Transistor (DEPFET) is a semiconductor
detector concept which combines detection and amplification within one
device. Due to it's low input capacitance the intrinsic S/N is very
high. In addition it can be operated at very low power. Both features
open a lot of application fields in High Energy Physics and astronomy.
Optimization of technology and design need elaborated simulation tools.
The talk covers design and simulation of a DEPFET detector array for the
Belle II experiment at KEK in Japan.
Priv.-Doz. Dr. A. Glitzky (WIAS, FG1)
Existence of bounded steady state solutions to spin-polarized
drift-diffusion systems
We study a stationary spin-polarized drift-diffusion model for
semiconductor spintronic devices. This coupled system of continuity
equations and a Poisson equation with mixed boundary conditions has to
be considered in heterostructures. In 3D we prove existence and
boundedness of steady states. If the Dirichlet conditions are
compatible or nearly compatible with thermodynamic equilibrium the
solution is unique. The same properties are fulfilled for a space
discretized version of the problem: Using a Scharfetter-Gummel scheme
on 3D boundary conforming Delaunay grids, the existence, boundedness
and, for small applied voltages, the uniqueness of the discrete
solution is obtained.
Dr. H. Si (WIAS, FG3)
On 3D Delaunay mesh generation
This talk discusses the problem of generating 3D tetrahedral meshes
whose elements satisfy the Delaunay criterion. It is motivated by
finite volume algorithms for the numerical solution of PDEs.
In this talk, we present a practical technique for solving this
problem. Our method first recovers domain boundary by constructing a
constrained Delaunay tetrahedralization, then it improves the mesh
quality by a variant of Delaunay refinement algorithm. An
implementation is available in the software TetGen.
H.-J. Diersch (DHI-WASY GmbH)
Computational aspects in porous media problems Name
Flow, mass and heat transport through porous media encounter in many
branches of engineering and science. Of particular concern are those
processes occurring beneath the surface of the earth's ground, that
means subsurface flow and transport in geologic media with their
complexity and uncertainty. Men are also looking for new technologies
of exploiting geothermal energy and storing fluids in reservoirs.
Industries are developing new materials with improved properties for
which a greater understanding of flow and energy transport is
required. There is a wide spectrum in porous media flow modeling
ranging from multi-phase flow via chemical reaction systems, fracture
flow modeling, deformation processes to different numerical
approaches.
DHI-WASY is the developer of the commercial finite-element simulator
FEFLOW. We briefly characterize the status quo regarding the
capabilities available in FEFLOW, numerical features and inherent
software technologies. The present paper will discuss computational
aspects along three chosen fields of porous media problems for 2D and
3D applications.
(1) Multi-diffusive fingering convection processes: New powerful
features of the simulator are presented. It covers extensions to
multi-species reactive transport equations, species and temperature
related multiple density expansion parameters and multiple fluid
viscosity relations.
(2) Unsaturated hysteretic flow in absorbent swelling porous media:
The flow and deformation processes in swelling porous media are
modeled for absorbent hygiene products (e.g., diapers, wipes,
papers). The governing modeling equations are strongly nonlinear and
requires advanced numerical strategies for solving. Mesh movement and
mesh-refinement strategies are incorporated. Spline approximations are
used for better and more flexible descriptions of experimental data
and measured relations.
(3) Modeling of borehole heat exchanger (BHE) systems: We briefly
discuss fundamental equations for BHE systems and their finite-element
representations. Improved relationships for thermal resistances of BHE
has been developed. The numerical solution of the final 3D problems is
performed via a widely non-sequential (essentially non-iterative)
coupling strategy for the BHE and porous medium
discretization. Practical application to a borehole thermal energy
store (BTES) consisting of 80 BHE is given for the real-site BTES
Crailsheim, Germany. The simulations are controlled by the
specifically developed FEFLOW-TRNSYS coupling module. Scenarios
indicate the effect of the groundwater flow regime on efficiency and
reliability of the subsurface heat storage system.
Prof. H. Gajewski (WIAS)
On a variational approach for domain separation
We are interested in algorithms for constructing surfaces Γ of
possibly small measure that separate a given domain Ω into two
regions of equal measure. Using the integral formula for the total
gradient variation, we show that such separators can be constructed
approximatively by means of sign changing eigenfunctions of the
p-Laplacians, p → 1 , under homogeneous Neumann boundary
conditions. These eigenfunctions turn out to be limits of steepest
descent methods applied to suitable norm quotients.