# CE Seminar

Together with the Computational Engineering Research Center of the TU Darmstadt a joint seminar with interesting talks in the field of CE is organized in every semester. If you are interested in these seminars and would like to receive invitations please subscribe for the corresponding mailing list.

## 2017

## Waveform Relaxation on Circuit Differential-Algebraic Equations: A Topological Convergence Criterion

### Jonas Pade, M.Sc., Humboldt-University of Berlin

**18 Dec 2017, 16:15–17:45; Location: S2|17-103 **

While waveform relaxation methods converge for coupled ODE systems, they may suffer from instabilities for coupled differential-algebraic equations (DAEs). Whereas several convergence criteria are known for index-1 DAEs, we present here a convergence criterion for quasilinear index-2 DAEs for the first time. It applies to systems of coupled ODE/circuit systems. Such systems appear in co-simulation approaches for coupled EM/circuit simulation. Under certain coupling assumptions, we can give a topological interpretation to the convergence criterion.

## Mobile technology for production-ready high-performance computing systems: The path of the Mont-Blanc project

### Filipo Mantovani, Barcelona Supercomputing Center

**23 Oct 2017, 16:30–18:00; Location: S2|02-C110**

Since 2011 the EU Mont-Blanc project pushes the development of ARM-based compute platforms, following the vision of leveraging the fast growing market of mobile technology for performing scientific computation. The process started almost 5 years ago with the development of prototypes based on Android dev-kits and is now evolving beyond the research project, towards commercial computational platforms based not only on mobile SoCs, but also on server and HPC technology.

In this talk we will introduce the ARM-based prototypes deployed within the time frame of the project as well as the experience gained porting system software, performance analysis and debugging tools, and real HPC scientific applications. The goal of the talk is to give an overview of ARM based scientific computing from the Mont-Blanc perspective, supported by experience, lesson learned and test results.

## Coupled Simulations of Electric Arcs for Switching Devices with MpCCI and ANSYS

### Pascal Bayrasy, Fraunhofer-Institute for Algorithms and Scientific Computing SCAI, St. Augustin, Germany

**6 Oct 2017, 14:00–15:30; Location: S2|17-103**

## Bunch trains in the Linac Coherent Light Source

### Dr. Karl Bane, SLAC National Accelerator Laboratory, USA

**11 Sep 2017, 16:30–18:00; Location: S2|17-103**

The Linac Coherent Light Source is a linac-based, X-ray free electron laser (FEL) at the SLAC National Accelerator Laboratory. It accelerates single bunches up to 4--13 GeV before sending them to the undulator for lasing. A project is underway to run the LCLS bunches in trains, with up to initially four and later eight bunches in a train. Bunch trains will allow the users to perform pump-pump-probe experiments.

However, running bunch trains in the linacs of the LCLS can excite transverse wakefields, resulting in the beam break-up instability and projected emittance growth. In this talk we estimate the severity of the problem and investigate mitigation methods.

## Towards pressure-robust mixed methods for the incompressible Navier-Stokes equations

### Dr. Alexander Linke, Weierstrass Institute for Applied Analysis and Stochastics, Berlin

**20 Jul 2017, 17:00–18:30; Location: S4|10-1**

For more than thirty years it was thought that the efficient construction of pressure-robust mixed methods for the incompressible Navier-Stokes equations, whose velocity error is pressure-independent, was practically impossible. However, a novel, quite universal construction approach shows that it is indeed rather easy to construct pressure-robust mixed methods. The approach repairs a certain (L2-)orthogonality between gradient fields and discretely divergence-free test functions, and works for families of arbitrary-order mixed finite element methods, arbitrary-order discontinuous Galerkin methods, and finite volume methods. Novel benchmarks for the incompressible Navier-Stokes equations show that the approach promises significant speedups in computational practice, whenever the continuous pressure is complicated.

## Efficient calculation of spectral bounds for Hessian matrices on hyperrectangles

### Prof. Dr.-Ing. Martin Mönnigmann, Ruhr-Universität Bochum

**18 Jul 2017, 17:00–18:30; Location: S4|10-1 **

The presentation summarizes progress made over the past few years in the calculation of spectral bounds of interval Hessian matrices. Spectral bounds of this type play an important role in global optimization algorithms. They can be used, for example, to detect if a given optimization problem is convex, or to generate a convex relaxation in case it is not. In technical terms, the problem is the following: Find, in a computationally efficient way, bounds on the eigenvalues of all Hessian matrices in a Hessian matrix set for \(\{\nabla^2\Phi(x)\vee x\in S\}\) a \(C^2\)-function \(\Phi:U\subset R^n\rightarrow R\ \) on a hyperrectangle \(S\subset U\). The new method differs from existing ones in that it deliberately does not use any interval matrices. As a result, it exhibits two interesting features: The new method requires only \(\mathcal{O}(n) N(\Phi)\) operations (where \(N(\Phi)\) refers to the number of operations necessary to evaluate \(\Phi\) at a given point). Secondly, for some functions \(\Phi\), the new method results in tighter eigenvalue bounds than the tightest theoretically possible bounds for the interval Hessian matrix. This is surprising, since the fastest method for calculating the tightest possible eigenvalue bounds for the interval Hessian requires \(\mathcal{O}(2^n)\) operations, in contrast to the \(\mathcal{O}(n)N(\Phi)\) operations required here. In this sense, the proposed method belongs to a much more attractive complexity class and it sometimes results in better bounds than one of the best known methods.

## Bayesian subset simulation

### Prof. Julien Bect, CentraleSupélec, France

**17 Jul 2017, 16:15–17:45; Location: S2|17-103**

The numerical computation of small probabilities of failure for a system is a notoriously difficult problem, and the Subset Simulation algorithm of Au & Beck (Prob. Eng . Mech., 2001) has become one of the most popular method to solve it. Subset simulation has been shown to provide significant savings in the number of evaluations of the performance functions, with respect to a brute-force Monte Carlo approach. The number of evaluations remains quite high, however, for many practical applications where the performance function f is provided by an expensive-to-evaluate computer model.

This seminar will focus a stochastic algorithm, called Bayesian subset simulation (BSS), that uses a blend of ideas from the subset simulation approach and from the sequential design of computer experiments based on Gaussian process (GP) models. A key idea, as in the subset simulation algorithm, is to estimate the probabilities of a sequence of excursion sets of f above intermediate thresholds, using a sequential Monte Carlo (SMC) approach. A GP prior on f is used to define the sequence of densities targeted by the SMC algorithm, and drive the selection of evaluation points of f to estimate the intermediate probabilities. Adaptive procedures are proposed to determine the intermediate thresholds and the number of evaluations to be carried out at each stage of the algorithm.

We will discuss the main ideas behind the method and present a free software implementation for Matlab/Octave (using the STK toolbox) that makes it easy for you to try BSS on your own problems.

Main ref: Julien Bect, Ling Li and Emmanuel Vazquez. Bayesian subset simulation. https://arxiv.org/abs/1601.02557. SIAM Journal on Uncertainty Quantification (accepted).

## Numerical approximation of charged particle motion in magnetic quadrupoles

### Abele Simona and Prof. Luca Bonaventura, Politecnico di Milano

**13 Mar 2017, 16:15–17:45; Location: S2|17-103**

We present some attempts at improving the accuracy and efficiency of the reference numerical methods presently employed to model charged particle motion in magnetic quadrupoles. We first introduce the standard technique used to reconstruct the quadrupole magnetic vector potential, based on general solutions of the Laplace equation in cylindric coordinates. We then compare various time integration methods, in order to assess how a reduction of the (large) computational cost involved could be achieved.

In particular, we consider some of the methods used in the accelerator physics community, based on a Lie algebra approach, along with other symplectic methods and non symplectic ones. The results show that higher order methods could provide a more efficient alternative to the lower order approximations presently employed.

This is joint work with the group of Dr. Barbara Dalena of CEA (Commissariat à l'énergie atomique et aux énergies alternatives), Saclay, France.

## A Study of Dynamic Loop Scheduling Algorithms on Modern Manycore Architectures

### Prof. Dr. Florina Ciorba, Universität Basel

**20 Feb 2017, 16:30–18:00; Location: S2|02-C110**

High performance computing (HPC) systems are nowadays large, consist of diverse components, and exhibit massive parallelism. Exploiting the available hardware parallelism is notoriously challenging, in part due to difficulty in exposing and expressing parallelism in the computational applications. Loops are considered the main source of parallelism in most scientific computational applications. Dynamic loop scheduling (DLS) algorithms are used to parallelize scientific applications and achieve optimized execution performance.

In this talk we examine the effect of two different parallel programming models (multithreaded and message-passing) on the performance of four well-known DLS algorithms: static chunking (STATIC), self service (SS), guided self scheduling (GSS), and factoring (FAC). Three different parallel programming libraries (Pthreads, OpenMP, and OpenMPI) are exploited for parallelizing matrix multiplication and adjoint convolution on the Intel Xeon Phi Knights Landing (KNL) architecture.

## NVM and HPC – Is there a use case?

### Prof. Dr. André Brinkmann, Johannes Gutenberg Universität Mainz

**13 Feb 2017, 16:30–18:00; Location: S2|02-C120**

Parallel file systems (PFS) have been developed to support typical HPC applications, where most file system operations include reading or writing huge chunks of data, which can be easily striped over the storage backend. New data-intensive applications will change these patterns and small random I/Os will occur more frequently in HPC, missing the sweet-spot of today’s parallel file systems. At the same time, supercomputers start to include node-local flash-storage or NVRAM, offering the opportunity to decouple application I/O during the application’s runtime from the PFS, pushing down the PFS one step in the storage hierarchy. This talk will discuss the impact of new memory and storage technologies on HPC and requirements and design-considerations to dynamically support the usage of node-local storage within application workflows.

## Parallel-in-Time Integration with PFASST

### Dr. Robert Speck, Forschungszentrum Juelich GmbH

**30 Jan 2017, 16:15–17:45; Location: S2|17-103**

For time-dependent partial differential equations, parallel-in-time integration using the “parallel full approximation scheme in space and time” (PFASST) is a promising way to accelerate existing space-parallel approaches beyond their scaling limits. Inspired by the classical Parareal method and non-linear multigrid ideas, PFASST allows to integrate multiple time-steps simultaneously using “multi-level spectral deferred corrections” (MLSDC) with different coarsening strategies in space and time. In numerous studies, this approach has been successfully coupled to space-parallel solvers which use finite differences, spectral methods or even particles for discretization in space. In this talk, we highlight in particular the interweaving of PFASST with a parallel multigrid solver in space and show extreme-scale benchmarks on up to 448K cores of the IBM Blue Gene/Q installation JUQUEEN. We then formulate PFASST itself as a specialized FAS multigrid method to provide a much easier access to the mathematical analysis and algorithmic optimization of this approach. In addition, we discuss possible extensions and sketch further research directions for the future of PFASST.

## Numerical Simulations of Insect Flight in Heavy Turbulence

### Thomas Engels, Ph.D., TU Berlin / École Normale Supérieure Paris

**12 Jan 2017, 17:00–18:30; Location: S4|10-1**

The aerodynamics of flying insects is a field of research currently receiving much attention from a broad, interdisciplinary community of researchers. Their motivation ranges from the design of bio-inspired robots to fundamental questions in biology and fluid mechanics. Insects are known to produce a surprising amount of force, using comparatively small wings without aerodynamic profile, by exploiting a vortex system formed in every stroke at the leading edge of their wings. Using high-resolution numerical simulations of a tethered model bumblebee in forward flight, we investigate whether environmental turbulence, ubiquitous in natural habitats, can destroy this vortex system crucial to the animals flight performance. Our numerical bumblebee is a prototype for medium sized insects (Reynolds number 2000), and first considered in laminar flow. Subsequently, we superimpose homogeneous isotropic turbulent fluctuations to the uniform inflow. Despite tremendous variation in turbulence intensity, between 17% and 99% with respect to the mean flow, we do not find significant changes in cycle-averaged aerodynamic forces, moments, or flight power when averaged over realizations, compared to laminar inflow conditions. The variance of aerodynamic measures, however, significantly increases with increasing turbulence intensity, which may explain flight instabilities observed in freely flying bees.