Scope Program Committee Registration Travel & Locations


All times are given in Central European Summer Time (CEST). An additional live stream will be available. Please contact Mr. Johannes Gebert ( if you would like to participate.

Monday, May 23rd, 2022

10:00 - 10:15 Welcome & Introduction
Michael Resch, HLRS, University of Stuttgart
10:15 - 10:45 Reaggregation of Disaggregation:
A Smart Approach to the Optimized Architecture and Platform Design

Hiroaki Kobayashi, Tohoku University
10:45 - 11:15 Combinatorial Clustering for a Material Informatics Application using Aurora Vector Annealing
Kazuhiko Komatsu, Tohoku University
Abstract Due to the recent advancement of data science, such as machine learning and big-data analysis, the approach using data science techniques has attracted attention even to develop new materials, called material informatics. In material informatics, clustering is one of the essential data processing techniques to understand thermophysical properties. To improve clustering accuracy, this presentation gives an Ising-based clustering method using aurora annealing machine for a material informatics application.
11:15 - 11:45 HPC Refactoring Catalog : Updates
Ryusuke Egawa, Tokyo Denki University
Abstract Aiming to support smooth code migration and optimization among HPC systems, HPC Refactoring, a database of system-aware code optimization patterns, was designed in 2015. This talk introduces an overview and updates/future plans for improving the HPC refactoring catalog.
11:45 - 13:15 Lunch
13:15 - 13:45 Concept of a File Tracing Mechanism for Research Data Management in High Performance Computing System
Yuta Namiki, Takeo Hosomi, AKihiro Yamashita, Susumu Date, Joint Research Laboratory for Integrated Infrastructure of High Performance Computing and Data Analysis, Cybermedia Center, Osaka University
Abstract Research data management has come to take a role of great importance for reproducible and reusable research. In the recent academic research scene, researchers and scientists leverage HPC systems as a base for processing and analyzing a large amount of data through numerical analysis and computer simulations. Despite that, however, the data produced and analyzed on HPC systems are not managed due to the lack of a functionality that allows us to understand how data is produced and processed in the system. From the perspective, by focusing on the lineage of data, which shows the origins and history of data for assuring that the data is produced on HPC systems. we have prototyped a mechanism that generates the lineage by tracing file access operations of user programs in the kernel. In this presentation, we report the interim result we have achieved so far with future issues.
13:45 - 14:15 Mitigation of Aeroacoustic Noise based on Simulations on HPC Systems
Matthias Meinke, Ansgar Niemoeller, Miro Gondrum, Moritz Waldmann, and Wolfgang Schroeder, AIA, RWTH Aachen
Abstract The numerical simulation of aeroacoustic sound is important for an improved understanding of noise generation mechanisms and the design of noise mitigation strategies. In this paper, the performance of two direct coupled two-step CFD/CAA methods implemented on HPC hardware are discussed. For the flow field either a finite-volume method for the solution of the Navier-Stokes equations or a lattice Boltzmann method is coupled to a discontinuous Galerkin method for the solution of the acoustic pertubation equations. The coupling takes advantage of a joint Cartesian mesh allowing for the exchange of the acoustic sources without MPI communication. An immersed boundary treatment of the acoustic scatttering from solid bodies by a novel solid wall formulation is implemented and validated in the DG method. Results for the case of a spinning vortex pair and the low Reynolds number unsteady flow around a circular cylinder show that a solution with comparable accuracy is obtained for the two direct hybrid methods when using identical mesh resolution. Finally, results of a large scale application, i.e., the noise prediction for a nose landing gear are presented.
14:15 - 14:45 Management of data flows between Cloud, HPC and IoT/Edge
Kamil Tokamov, HLRS, University of Stuttgart
Abstract The components of heterogeneous applications are deployed across various execution platforms and utilise the capabilities of the platforms. As such, one component can utilise HPC resources for better performance in batch computation, while another - Cloud resources, for better scalability and elasticity. Furthermore, this is also a possibility for processing on Edge devices. The usage of such a hybrid setup, where dependent components of the applications are deployed across various platforms, might require flexible and adaptive data transfers from one platform to another. This work presents a data management framework, based on the Apache NiFi dataflow management system and developed in the scope of the SODALITE EU project. This framework enables scalable data transfer between any of GridFTP (a file transfer protocol dominant in HPC), HTTP, S3-compatible and data streaming (such as MQTT) endpoints.
14:45 - 15:15 Break
15:15 - 15:45
Sabine Roller, DLR (Deutsches Zentrum für Luft- und Raumfahrt e.V.)
Abstract The abstract will be provided soon.
15:45 - 16:15 Integration of parallel HDF5 I/O in a large scale computational fluid dynamics solver
Tobias Gibis, University of Stuttgart
Abstract As the number of cores in massively parallel computer systems increases, I/O strategies must be adapted as to not present bottlenecks. The "read and write one file per core" strategy, while efficient for smaller earlier computational architectures, leads to an unmanageable amount of files and is poorly suited for Lustre filesystems. With the introduction of Hawk at HLRS, an urgent need had arisen to develop a new framework based on MPI-I/O in which all cores write simultaneously to a common file. To this end, a new I/O framework based on HDF5 was implemented in the IAG in-house CFD code NS3D. The associated talk will discuss selected design decisions and challenges encountered, escpecially regarding adequate I/O performance.
16:15 - 16:45 High Performance Object-Oriented Data Processing Workflows for Researchers and Scientists
Jason Appelbaum , University of Stuttgart
Abstract As computing capacity increases, datasets generated by HPC applications grow in size as well. The researchers and scientists who use such datasets for their work require scalability and efficiency in their data-processing workflows, but still prioritize utility and practicality. Typically such researchers are self-taught, intermediate-level programmers working collaboratively with others, in which case object-oriented languages such as python offer a greatly reduced barrier to entry and improved code maintainability within their research groups. Research progress is accelerated by flexible, easy access to large datasets for comparison and testing. The merits of a high-performance yet flexible approach utilizing HDF5 and MPI, wrapped by h5py and mpi4py in python, will be discussed. Examples taking advantage of collective I/O and parallel processing for the analysis of direct numerical simulation datasets will be presented, along with performance metrics. Additionally, techniques for 'big data' visualization using Paraview, HDF5 and XDMF will be showcased.
18:30 Dinner (Registration closed)

Tuesday, May 24th, 2022

09:00 - 09:45 Keynote
Jack Dongarra , University of Tennessee
09:45 - 10:15 A ML-Based Approach to Automatic Selection of Compiler and its Option Flags
Hiroyuki Takizawa, Tohoku University
Abstract Today, one HPC platform could have multiple compilers, each of which provides a lot of option flags. Those compilers have different optimization capabilities, and target at even different processors on a heterogeneous computing system such as NEC SX-Aurora TSUBASA. Thus, it could be challenging to select an appropriate build configuration such as the best available compiler and its option flags for each application code. In this talk, I will introduce our ongoting work to use machine learning for predicting an appropriate build configuration from performance counter values.
10:15 - 11:00 Break
11:00 - 11:30 Speeding up k-nearest neighbors search with space-filling curve.
Masashi Kotera, Sourav Saha, Takuya Araki, NEC Corporation
Abstract k-nearest neighbors search (k-NN) is a useful algorithm that can be used for classification and regression, but naive k-NN is slow because it requires scanning all the training data at prediction time. There is a method of solving this problem for low dimensional data, such as dividing the search space Like kd-tree. However, algorithms using tree structures are difficult to vectorize because they require recursion. In this talk, we talk about implementing the speed-up method for k-NN using z-curve, a kind of space-filling curve; the calculation of z-curve is easy to vectorize, and z-curve enables efficient range search that can be used to implement k-NN.
11:30 - 12:00 Heterogeneous Computing with SX-Aurora TSUBASA Vector Engine
Ryota Ishihara, NEC Corporation
Abstract SX-Aurora TSUBASA supports a variety of execution models in heterogeneous environments including various computational resources such as GPU and x86. User can select appropriate computational resources according to characteristics of each of applications in the executions. In this session, we will introduce the functions provided in each execution model and how to use them.
12:00 - 13:30 Lunch
13:30 - 14:00 Trends
Michael Resch, HLRS, University of Stuttgart
14:00 - 14:30 Searching a roadmap to solve partial differential equations with quantum machine learning
Markus Mieth, Pia Siegl, DLR (German Aerospace Center)
Abstract Our research aims to evaluate the potential of quantum computing to solve partial differential equations (PDEs) in the context of aerospace engineering. Established algorithms and methods for PDEs often rely on discretization in time and space. For a reasonable accuracy they come with high computational costs in terms of time and memory space. Machine learning approaches are studied as an alternative. One promising concept is the physical informed neural network (PINN) [1]. Here, the PDE is directly included into the loss function such that no data or only a limited amount is needed for training. In our approach, we exchange the classical neural network of the PINN with a trainable quantum circuit, while the optimization still runs on a classical computer. While we can already approximate simple functions successfully with known strategies [2, 3], more complex PDEs are hard to solve. Our work focuses on the search of problem-oriented quantum circuits and data encoding strategies, which increase the expressibility of the quantum model and allow for the approximation of more complex functions.
14:30 - 15:00 Prediction of Bio-Hybrid Fuel Injection and Mixture Formation in an Internal Combustion Engine
Tim Wegmann, Matthias Meinke, Wolfgang Schroeder, AIA, RWTH Aachen
Abstract For an efficient, stable and low emission combustion of novel e-fuels in piston engines, the fuel distribution at start of ignition plays a crucial rule. The injection system and fuel properties define the initial fuel vapor distribution. The subsequent fuel-air mixing depends on the convection, turbulence intensities, and the formation and break-up of large-scale flow structures, in particular the tumble vortex. Large-eddy simulations (LES) with high mesh resolution are necessary to accurately predict all involved scales for the mixing process. In this study, numerical analyses of the liquid fuel injection and the fuel-air mixing in a piston engine are performed. LES are conducted using a hierachical unstructured Cartesian mesh method with an efficient four-way coupling of the spray droplets with the gas phase. Due to the large number of spray droplets, a Lagrangian Particle Tracking (LPT) algorithm is used to accurately predict the liquid spray propagation and evaporation. The spray model is based on a KHRT-breakup formulation. The Navier-Stokes equations are solved for compressible flow using a finite-volume method, where boundary surfaces are represented by a conservative cut-cell method. The hierarchical Cartesian mesh ensures efficient use of high performance computing platforms through solution adaptive refinement and dynamic load balancing.
15:00 - 15:30 Break
15:30 - 16:00 Quantum machine learning for data analysis
Li Zhong, HLRS, University of Stuttgart
Abstract Fault-tolerant quantum computers have been proven to be able to improve machine learning through speed-ups in computation orimproved model scalability. Therefore, research at the junction of the two fields has garnered an increasing amount of interest, which has led to the rapid development of quantum deep learning and quantum-inspired deep learning techniques. In thiswork, we will demonstrate how quantum computers and quantum algorithms can be leveraged for image processing through quantum-inspired deep neural networks.
16:00 Open End