KCSE Science Day

Wednesday 21 Nov 2012, 10:00 → 15:00 Europe/Stockholm

Erwin Laure (PDC - Center for High-Performance Computing), Michael Hanke (CSC - KTH School of Computer Science and Communication)

Invitation to the
KCSE Science Day on High-Performance Computing
Internationally known experts will present their views on history, current status and future of high-performance computing and its applications in natural and engineering sciences. An integral part of this event will be presentations by KCSE students, which are related to usage of high performance computing in their PhD studies.

Everybody is invited to attend. No registration necessary.

Invited Speakers: Thomas Ludwig, Director DKRZ, <it>The History of Parallel Computing</it> Jesper Träff, Technical University Vienna, <it>History and Development of MPI: the Message-Passing Interface</it> The aim of this event is four-fold. We intend to discuss topics of interest to members of KCSE and a broad audience, have high-level experts discussing recent developments, provide an internal platform for the students to present their results, and promote networking.

Important Dates: Abstract submission deadline: September 21st, 2012 Notification of acceptance: October 20, 2012 Michael Hanke, Director KCSE (Numerical Analysis) Erwin Laure, Chairman of Science Day (PDC) Zilvinas Rinkevicius, Director of Studies (Theoretical Chemistry) Philipp Schlatter (Mechanics) Lars Bergqvist (Material Science) Anatoly Belonoshko (Theoretical Physics)

09:30 → 09:40 Welcome

Speakers: Erwin Laure (PDC-HPC), Michael Hanke

09:40 → 10:00 Scaling Dalton, a molecular electronic structure program

Dalton is a molecular electronic structure program used for quantum mechanics (QM) and quantum mechanics/ molecular mechanics (QM/MM). It is specialized and has a leading position in calculation of molecular properties, having a large world-wide user community (over 2000 licenses issued). In this work, we present an extensive characterization and performance optimization of Dalton that increases its scalability and parallel efficiency.

Speaker: Xavi Aguilar

10:00 → 10:20 A new high order method for the accurate simulation of incompressible wall-bounded flows

The aim of the project is to develop an efficient high-order method for Direct Numerical Simulations and Large Eddy Simulations of wall-bounded incompressible flows with accurate fulfillment of the divergence-free condition. One of the main issues when numerically solving the Navier– Stokes equations is the absence of an evolution equation for the pressure. Over the years several methods have been developed to deal with this, among the most popular are writing the equations in vorticity-velocity form, and predictor-corrector methods like the pressure correction or fractional step method. In the former method the pressure is eliminated by taking the curl of the momentum equations, while in the latter method a Poisson equation for the pressure is derived by taking the divergence of the momentum equations. But in wall-bounded domains unclarity still remains about the choice of boundary conditions for the pressure. Most implementations of the pressure correction method use a staggered mesh which removes the need to solve for the pressure at the boundary, but it also can filter high frequencies in unintended ways and complicates implementation. Kleiser and Schumann developed the influence matrix method which calculates the boundary conditions for the pressure which ensure that the discrete diver- gence is numerically zero in the whole field. They used Fourier collocation in the stream- and spanwise direction and a Chebyshev-Tau method in wall-normal direction. We extend their method to include the use of compact finite differences on a collocated grid in wall-normal direction. Compact finite difference schemes have better resolution characteristics than the standard finite difference schemes, while maintaining flexibility in the choice of grid spacing and boundary conditions. The resulting Poisson equation for the pressure contains a full matrix in the left- hand side which is numerically expensive to solve. To lower the computational cost and memory requirements, we introduce a new method to solve the Poisson equation which replaces solving the second order full system by solving two first order banded equations. In summary, our method allows the solution of the incompressible Navier–Stokes equations with exactly zero divergence on collocated meshes, in a fully banded formulation.

Speaker: Peter Lenaers

10:20 → 10:40 Coffee

10:40 → 11:00 An algorithm for efficient evaluation of two-electron repulsion integrals over contracted Gaussian-type basis functions

A new algorithm for the evaluation of two-electron repulsion integrals optimized for high contraction degrees is derived. Both the segmented and general contraction versions of the algorithm show significant theoretical performance gains over the asymptotically fastest algorithms published in the literature so far. A preliminary implementation of the algorithm shows good agreement with the theoretical results and demonstrates substantial average speed-ups in the evaluation of two-electron repulsion integrals over commonly used basis sets with varying degrees of contraction with respect to a mature, highly optimized quantum chemical code.

Speaker: Rosal Sandberg

11:00 → 12:00 Keynote: History and Development of MPI: the Message-Passing Interface

MPI, the Message-Passing Interface, has, for better or worse, for a considerable number of years been the de facto programming interface for large scale scientific computing. In this talk we will examine some of the reasons why this has happened, focusing on some of the salient features of the MPI standard that has contributed to its usefulness and longevity. The talk will do so by also looking into the long and perhaps not so well-known history of the development of MPI, will and conclude with an outlook on the recently announced MPI 3.0 version of the standard.

Speaker: Jesper Träff (Technical University of Vienna)

Slides MPIHistory.pdf

12:00 → 13:30 Lunch

13:30 → 13:50 High performance adaptive finite element methods

The massive computational cost for resolving all turbulent scales makes a direct numerical simulation of the underlying Navier-Stokes equations impossible in most engineering applications. This work concerns the development of an adaptive finite element method that enables efficient computation of time resolved approximations for complex geometries with error control. We present efficient data structures and data decomposition methods for distributed unstructured tetrahedral meshes. Our work also concerns an efficient parallelization of local mesh refinement methods such as recursive longest edge bisection, and the development of an a priori predictive dynamic load balancing method, based on a weighted dual graph.We also address the challenges of emerging supercomputer architectures with the development of new hybrid parallel programming models, combining traditional message passing with lightweight one-sided communication. Our implementation has proven to be both general and efficient, scaling up to more than 12k cores.

Speaker: Niclas Jansson

13:50 → 14:50 Keynote: Parallel Computing - From 1757 until 2020 and beyond

High performance computing is an enabling factor for gaining of knowledge and insight in modern sciences. However, even 250 years ago mathematicians were confronted with the fact that a single computer was not powerful enough to complete complex calculations in acceptable time. The talk will give historical details about the history of parallel computing from the times of the Enlightenment until now. It will explain techniques and application fields and we will discuss the future of parallel computing.

Speaker: Thomas Ludwig (DKRZ)

14:50 → 15:10 Coffee

15:10 → 15:30 Efficient spike communication in the MUSIC multi-simulation framework

MUSIC is a standard API and software library allowing large- scale neuronal network simulators, or other applications, to exchange data within a parallel computer during runtime. In this talk we will show how the efficiency of point-to-point and collective communication algorithms depends on the topology of the connectivity between the applications. We will present benchmark results comparing the scaling of the two communication algorithms on high number of tasks, and will discuss challenges regarding scheduling of communication in MUSIC.

Speaker: Ekaterina Brocke

15:30 → 15:50 Parallelized calculation of vibrationally-resolved electronic spectra with DynaVib software

Vibrationally resolved electronic spectra can be simulated by calculating the overlap integrals of the vibrational wave functions of initial and final states. This method works fine with small molecules. However, it could be very time consuming for large molecule and high-performance computing is necessary. We have developed a software called DynaVib to calculate the vibrationally-resolved electronic spectra of polyatomic molecules. The parallelization of the code is achieved using OpenMP and MPI. The software can be used to study the optical absorption and emission spectra, photoelectron spectra, X- ray absorption spectra, etc.

Speaker: Guangjun Tian

15:50 → 16:10 Conclusions

Speakers: Erwin Laure (PDC-HPC), Michael Hanke