Loading Events

« All Events

  • This event has passed.

PRACEdays15

26 May 2015 - 28 May 2015

Details

Start:
26 May 2015
End:
28 May 2015
Event Category:

PRACEdays15: HPC Brings Industry and Academia Together

PRACEdays15, which ran from 26 to 28 May at the Aviva Stadium Conference Centre in Dublin, Ireland.

PRACEdays15 marked another successful edition of the annual PRACE Scientific and Industrial Conference. The Irish Centre for High End Computing hosted the conference and satellite events locally and, with this year’s theme being “Enable Science Foster Industry”, attracted close to 200 experts from academia and industry.

PRACEdays15_logo

PRACEdays15 marked another successful edition of the annual PRACE Scientific and Industrial Conference. The Irish Centre for High End Computing hosted the conference and satellite events locally and, with this year’s theme being “Enable Science Foster Industry”, attracted close to 200 experts from academia and industry.

The two satellite events focussing on Women in HPC and Exascale European projects proved very popular on 25 and 26 May and the EESI2 Project held its final conference on 28 and 29 May to round off the week. The PRACE User Forum invited all participants to an Open Session on Wednesday afternoon.

Finally, the PRACEdays15 Award for Best Poster was presented to Panchatcharam Mariappan for his poster entitled “GPU accelerated finite element method for radio frequency ablated cancer treatment“.

Highlights of the week included the keynote sessions given by well-known academic and industrial researchers from Europe and Asia, including the keynote speech by Masahiro Seki, President of the Research Organisation for Information Science and Technology (RIST), Japan. The European Commission added their vision to the programme with a presentation entitled “Implementing the European Strategy on High Performance Computing” There were six heavily subscribed parallel streams across various scientific and industrial themes, which proved very successful. The final panel discussion on 28 May entitled “Science and Industry: Partners for Innovation”, was moderated by Tom Wilkie of Scientific Computing World and brought together high level representatives from the European Commission, industry and academia.

Details of all aspects of the event (award, presentations, poster session, videos created during PRACEdays15 and the satellite events) can be found by following the links below.

The PRACEdays15 Conference Programme can be found below.

PD15Booklet

BOOKLET PDF DOWNLOAD

Welcome & Keynotes

This session took place Tuesday 26 May 2015.

Title: Welcome

Speaker

  • Sergi Girona, Chair of the PRACE Board of Directors

Title: Fossils, Physics and Fast Computers Unlocking a Virtual Past

Speaker

  • William Sellers, Faculty of Life Sciences, The University of Manchester

Abstract
The past is a fascinating place. It can tell us how we came to be like we are today, and it contains a huge range of bizarre creatures that are no longer alive. However since we do not yet have a suitable time machine, all our knowledge about the distant past comes from evidence preserved in the rocks around us. The most important source of evidence is from fossils. These are the preserved remains of animals and plants and they have been collected and studied by geologists for hundreds of years. However nowadays other disciplines want to get in on the fun. Engineers, physicists and computer scientists have developed techniques that help us find out more about fossil organisms. This talk will concentrate on what we can learn from studying the mechanics of fossil organisms using high performance computers. It will demonstrate the way early humans moved and what this tells us about the origins of moderns humans. It will also show how fast and how heavy the largest dinosaurs were and what this means about the way they lived. But most importantly it will explain how we can actually answer these questions scientifically and avoid some of the guesswork and exageration that has happened in the past.

PRESENTATION PDF DOWNLOAD

Title: Listening to Black Holes with Super Computers

Speaker

  • Sascha Husa, Relativity and Gravitation Group at the University of the Balearic Islands

Abstract
One century after Einstein’s theory of general relativity has revealed space and time as dynamical entities a new generation of gravitational wave detectors is starting operation, and the first detection of gravitational wave events is expected to push open a new window on the universe within the next 5 years. The experimental challenge to meet the tremendous sensitivity requirements of GW detectors is paralleled by the computational modelling challenge to accurately predict the complicated dependence of the wave signals on the masses and spins of the black holes. In this talk will report on a program to explore the gravitational wave signatures of coalescing black holes by solving the Einstein equations with high order finite difference mesh refinement methods for judiciously chosen cases, and the synthesis of analytical models from our numerical data and perturbative results. These models are already used to analyse data from gravitational wave detectors and will help to identify the first such signals ever to be observed.

PRESENTATION PDF DOWNLOAD

Title: HPC Simulation at EDF Enabling Energy Challenges

Speaker

  • Ange Caruso, Information Technologies Program Manager, Electricité de France R&D

Abstract
An industrial utility like EDF needs to better understand the behavior of energy infrastructures like power plants (nuclear, thermal, renewable,…), electrical networks, but also energy management. The objective is to increase safety, performance, lifetime, and optimize processes. To reach these goals, it is necessary to better understand various phenomena met inside the infrastructures, for example: nuclear components (containment building, PWR vessel, steam generator, fuel rods), networks (electrical grids) or energy management (quality of electricity), in order to win margins. This is done using various numerical softwares developed at EDF R&D. The use of HPC simulation allows new approaches and new perspectives. Some applications will be shown.

PRESENTATION PDF DOWNLOAD

Welcome & Keynotes

This session took place Wednesday 27 May 2015.

Title: Welcome from the Local Host

Speaker

  • Jean-Christophe (JC) Desplat, Director of the Irish Centre for High-End Computing (ICHEC)

PRESENTATION PDF DOWNLOAD

Title: Opening Address

Speaker

  • Sanzio Bassini, Chair of the PRACE Council

PRESENTATION PDF DOWNLOAD

Title: Present Status of RIST in Promotion of High Performance Computing Infrastructure in Japan

Speaker

  • Masahiro Seki, President of RIST, Japan

Abstract
High Performance Computing Infrastructure (HPCI) has been established in Japan as a platform for the integrated use of high performance computer systems including the K computer. HPCI currently integrates 12 systems to provide hierarchical capability, with the K computer as the flagship. Other supercomputers serve as the second layer systems, which play various unique roles. All the computer systems are connected via a high speed network and are operated with the same policy to realize common operational features such as single-sign-on. The mission of RIST includes: (1) call for proposals, (2) screening and awarding and (3) user support. Roughly speaking, RIST in Japan is like PRACE in Europe. In his presentation, he will describe the evaluation process including recent results, supporting activities for shared use, promotion activities for industrial use and publication management.

PRESENTATION PDF DOWNLOAD

Title: Towards Exascale: The Growing Pains of Industry Strength CAE Software

Speaker

  • Lee Margetts, Research Computing Services, The University of Manchester

Abstract
In the Exascale community, there is some concern that commercial computer aided engineering (CAE) software will not be ready to take advantage of Exascale systems when they eventually come online. This talk will consider this issue from three perspectives: (i) Industry end users whose business will benefit from early access to Exascale computers; (ii) Independent software vendors who develop and market engineering software and (iii) Open source software initiatives led by Universities and government laboratories. Each of these stakeholders has a unique set of needs and motivational drivers that, if linked together in a simple and elegant way, can lead to the development, use and exploitation of CAE software on Exascale systems. The Lee Margetts will draw upon academic experience as leader of an open source software project and business insight through roles at NAFEMS and PRACE to set out a possible roadmap towards Exascale CAE.

PRESENTATION PDF DOWNLOAD

Title: Implementing the European Strategy on High Performance Computing

Speaker

  • Augusto Burgueño Arjona, European Commission, DG Communications Networks, Content and Technology

Abstract
HPC is a strategic tool to transform big scientific, industrial and societal challenges into innovation and business opportunities. HPC is essential for modern scientific advances (e.g. understanding the human brain or climate change) as well as for industry to innovate in products and services. “Traditional” areas like manufacturing, oil&gas, pharmaceutical industry etc. consider HPC indispensable for innovation, but also emerging applications (like smart cities, personalized medicine, or cosmetics) benefit from the use of HPC and its convergence with Big Data and clouds. The most advanced countries in the world recognise this strategic role of HPC and have announced ambitious plans for building exascale technology and deploying state-of-the-art supercomputers in the following years. Europe has the technological know-how and market size to play a leading role in all areas: HPC technologies and systems, services and applications. The European HPC Strategy in Horizon 2020 combines three elements in an integrated and synergetic way: (a) developing the next generations of HPC towards exascale; (b) providing access to the best facilities and services for both industry and academia; and (c) achieving excellence in applications. The Commission has taken several ambitious steps to support this strategy, such as the establishment of a contractual Public Private Partnership (PPP) on HPC. Support to the HPC Strategy is expected to continue in the future Horizon 2020 work programmes.

PRESENTATION PDF DOWNLOAD

Scientific Sessions

This session took place Wednesday 27 May 2015.

EUROPEAN RESEARCH COUNCIL PROJECTS

Chair: Kenneth Ruud, Artic University of Norway

Title: Computational Challenges in Skeletal Tissue Engineering

Speaker

  • Liesbet Geris, Biomechanics Research Unit, University of Liège

Abstract
Tissue engineering (TE), the interdisciplinary field combining biomedical and engineering sciences in the search for functional man-made organ replacements, has key issues with the quantity and quality of the generated products. Protocols followed in the lab are mainly trial and error based, requiring a huge amount of manual interventions and lacking clear early time-point quality criteria to guide the process. As a result, these processes are very hard to scale up to industrial production levels. In many engineering sectors, in silico modeling is used as an inherent part of the R&D process. In this talk I will discuss a number of (compute intensive) examples demonstrating the contribution of in silico modeling to the bone tissue engineering process. A first example that will be discussed is the simulation of bioreactor processes. Currently, only a limited number of online read-outs is available which can be used to monitor and control the biological processes taking place inside the bioreactor. We developed a computational model of neotissue growth inside the bioreactor that, in combination with the experimental read-outs, allow for a quantification of the processes taking place inside the bioreactor. Scaffold geometry (curvature-based growth), fluid flow (Brinkman equation) and nutrient supply were simulated to affect the growth rate of the neotissue. The model captured the experimentally observed growth patterns qualitatively and quantitatively. Additionally, the model was able to calculate the micro-environmental cues (mechanical and nutrient-related) that cells experience both at the neotissue-free flow interface and inside the neotissue. The second example pertains to the assessment of the in vivo bone regeneration process. As normal fractures lead to successful healing in 90-95% of the cases, people in need of tissue engineering solutions often suffer from severe trauma, genetic disorders or comorbidities. One of these genetic disorders impacting the bone regeneration process is neurofibromatosis type I. Starting from an established computational model of bone regeneration, we examined the effect of the NF1 mutation on bone fracture healing by altering the parameter values of eight key factors which describe the aberrant cellular behavior of NF1 affected cells. We show that the computational model is able to predict the formation of a non-union and captures the wide variety of nonunion phenotypes observed in patients. A sensitivity analysis by “Design of Experiments” was used to identify the key contributors to the severity of the non-union

Title: HPC for Combustion Instabilities in Gas Turbines: The ERC INTECOCIS Project in Toulouse

Speaker

  • Gabriel Staffelbach, CERFACS

Abstract
Combustion accounts for 80% of the worlds energy. Ever progressing trends in design and research have and continue to yield spectacular changes for turbines, cars, rocket propulsion etc.. The joint ERC project INTECOCIS coordinated by CNRS (DR14 Midi Pyrénées) and lead by two research centres, Institut de Mécanique des Fluides de Toulouse and CERFACS aims at introducing recent progress in the field of High Performance Computing (HPC) for combustion simulation into studies of Combustion Instabilities. The project integrates experimental and industrial applications to build and validate tools built to predict combustion instabilities using modern high performance computing. This presentation will highlight the recent progress of the project.

PRESENTATION PDF DOWNLOAD

Title: Runtime Aware Architectures

Speaker

  • Mateo Valero, Director at Barcelona Supercomputing Center (BSC)

Abstract
In the last few years, the traditional ways to keep the increase of hardware performance to the rate predicted by the Moore’s Law have vanished. When uni-cores were the norm, hardware design was decoupled from the software stack thanks to a well defined Instruction Set Architecture (ISA). This simple interface allowed developing applications without worrying too much about the underlying hardware, while hardware designers were able to aggressively exploit instruction-level parallelism (ILP) in superscalar processors. With the eruption of multi-cores and parallel applications, this simple interface started to leak. As a consequence, the role of decoupling again applications from the hardware was moved to the runtime system. Efficiently using the underlying hardware from this runtime without exposing its complexities to the application has been the target of very active and prolific research in the last years. Current multi-cores are designed as simple symmetric multiprocessors (SMP) on a chip. However, we believe that this is not enough to overcome all the problems that multi-cores already have to face. It is our position that the runtime has to drive the design of future multicores to overcome the restrictions in terms of power, memory, programmability and resilience that multi-cores have. In this talk, we introduce a first approach towards a Runtime-Aware Architecture (RAA), a massively parallel architecture designed from the runtime’s perspective.

HOT LATTICE QUANTUM CHROMODYNAMICS

Chair: Sándor Katz, Institute for Theoretical Physics, Eötvös Loránd University, Budapest

Title: Simulation of Strongly Interacting Matter from Low to High Temperatures

Speaker

  • Stefan Krieg, Forschungszentrum Jülich, Germany

Abstract
The rapid transition from the quark-gluon-plasma ‘phase’ to the hadronic phase in the early universe and the QCD phase diagram are subjects of intense study in present heavy-ion experiments (LHC@CERN, RHIC@BNL, and the upcoming FAIR@GSI). This transition can be studied in a systematic way in Lattice QCD. We report on a continuum extrapolated result for the equation of state (EoS) of QCD with and without dynamical charm degree of freedom. With these results, we will be able to close the gap between the low temperature region, which can be described by the hadron resonance gas model, and the high temperature region, which can be described by (hard thermal loop) perturbation theory. For all our final results, the systematics are controlled, quark masses are set to their physical values, and the continuum limit is taken using at least three lattice spacings.

Title: From Quark Number Fluctuations to the QCD Phase Diagram

Speaker

  • Christian Schmidt, University of Bielefeld

Abstract
For the first few micro-seconds after the Big Bang, the universe was filled with a plasma of strongly interacting quarks and gluons, the QGP. Today, small droplets of QGP are created in heavy ion experiments. Recently, large experimental effort was undertaken to explore the phase diagram of QCD through a beam energy scan program of heavy ion collisions. We will review Lattice QCD computations of conserved charge fluctuations that are performed in order to make contact with these experiments. We then show that a comparison of fluctuations of conserved hadroinc charged from lattice QCD with experimental results allows to position the so called freeze-out points on the QCD phase diagram. Computational challenges, that boil down to a tremendous amount of inversion of large sparse matrices will be highlighted. Here, the method of choice is the iterative conjugate gradient solver, which in our case is bandwidth limited. On GPUs, e.g., we approach this problem by exposing more parallelism to the accelerator through inverting multiple right hand sides at the same time.

Title: Lattice Simulations of Strong Interactions in Background Fields

Speaker

  • Massimo D’Elia, Chariman, University of Pisa & INFN, Italy

Abstract
The study of strong interactions in the presence of external sources, such as electromagnetic background fields or chemical potentials, offers the possibility to investigate the properties of strongly interacting matter in unusual conditions, which may be relevant to many contexts, going from heavy ion experiments to the physics of the Early Universe. Due to the non-perturbative nature of the problem, numerical lattice simulations are the ideal tool to obtain answers based on the first principles of Quantum Chromodynamics (QCD), which is the theory of strong interactions. The last few years have seen a considerable and steady progress in the field. Because of the extremely high computational needs of the problem, this has been possible also due to a matching development in HPC infrastructures. In this talk I will review such progress, with a focus on results regarding the physics of QCD in strong magnetic fields.

PRESENTATION PDF DOWNLOAD

COMPUTATIONAL DYNAMICS

Chair: Petros Koumoutsakos, Computational Science Lab, ETH Zürich, Switzerland

Title: Computational Challenges of Fast Dynamics of Flows with Phase Interfaces for Biomedical Applications

Speaker

  • Nikolaus Adams, Lehrstuhl für Aerodynamik und Strömungsmechanik – TU München, Germany

Abstract
The simulation of two-phase flows with compressibility effects and turbulence is one of the current challenges for modern numerical models in predictive simulations. Different approaches promise the most efficient way to solution for different application scenarios. The interaction of phase interfaces with shock waves and the generation of shock waves by rapid phase change are essential flow phenomena for biomedical applications. In this talk we present recent developments in modeling and simulation of compressible flows with interfaces, address efficient computational approaches for interface tracking, multi-resolution approaches, and new physically motivated approaches for dynamic load balancing.

Title: High Order, Scale Resolving Modelling for High Reynolds Number Racing Car Aerodynamics

Speaker

  • Spencer Sherwin, Faculty of Engineering, Department of Aeronautics at Imperial College London

Abstract
The simulation of two-phase flows with compressibility effects and turbulence is one of the current challenges for modern numerical models in predictive simulations. Different approaches promise the most efficient way to solution for different application scenarios. The interaction of phase interfaces with shock waves and the generation of shock waves by rapid phase change are essential flow phenomena for biomedical applications. In this talk we present recent developments in modeling and simulation of compressible flows with interfaces, address efficient computational approaches for interface tracking, multi-resolution approaches, and new physically motivated approaches for dynamic load balancing.The use of computational tools in industrial flow simulations is well established. As engineering design continues to evolve and become ever more complex there is an increasing demand for more accurate transient flow simulations. It can, using existing methods, be extremely costly in computational terms to achieve sufficient accuracy in these simulations. Accordingly, advanced engineering industries, such as the Formula 1 (F1) industry, is looking to academia to develop the next generation of techniques which may provide a mechanism for more accurate simulations without excessive increases in cost. Currently, the most established methods for industrial flow simulations, including F1, are based upon the Reynolds Averaged Navier-Stokes (RANS) equations which are at the heart of most commercial codes. There is naturally an implicit assumption in this approach of a steady state solution. In practice, however, many industrial problems involve unsteady or transient flows which the RANS techniques are not well equipped to deal with. In order to therefore address increasing demand for more physical models in engineering design, commercial codes do include unsteady extensions such as URANS (Unsteady RANS), and Direct Eddy Simulation (DES). Unfortunately even on high performance computing facilities these types of computational models require significantly more execution time which, to date, has not been matched with a corresponding increase in accuracy of a level sufficient to justify this costs. Particularly when considering the computing restrictions the F1 rules impose on the race car design. Alternative high order transient simulation techniques using spectral/hp element discretisations have been developed within research and academic communities over the past few decades. These methods have generally been applied to more academic transient flow simulations with a significantly reduced level of turbulence modelling. As the industrial demand for transient simulations becomes greater and the computer “power per $” improves, alternative computational techniques such as high order spectral/hp element discretisations, not yet widely adopted by industry, are likely to provide a more cost effective tool from the perspective of computational time for a high level of accuracy. In this presentation we will outline the demands imposed on computational aerodynamics within the highly competitive F1 race car design and discuss the next generation of transient flow modelling that the industry is looking to impact on this design cycle.

PRESENTATION PDF DOWNLOAD

Title: Salvinia-Inspired Surfaces in Action

Speaker

  • Carlo Massimo Casciola, University of Rome “La Sapienza”

Abstract
Surfaces exhibiting extraordinary features exist in nature. A remarkable example is the Salvinia molesta. This water fern, due the presence of small hydrophilic patches on top of rough hydrophobic surfaces, is able to retain air pockets when submerged by stabilizing the resulting Cassie state against positive pressure fluctuations while, at the same time, preventing bubble detachment. A similar strategy is adopted by certain insects and spiders (e.g. Elmis Maugei and Dolomedes triton) to breath underwater, thanks to a stabilized air layer, the so-called plastron. However, since CWT is a rare event beyond reach of brute force computations, the mechanism of wetting remains elusive and it is still difficult, if not impossible, to predict the transition from the Cassie to the fully wet Wenzel state. Using specialized techniques, it has been recently demonstrated that molecular dynamics is indeed capable to describe the Cassie-Wenzel transition on a simple model system. However, going beyond this proof-of-concept simulations, with the goal of reproducing real hydrophobic coatings and the complex morphology of natural surfaces, requires to combine a smart theoretical approach with a boost in computational resources. We discuss here the results of massively parallel simulations on top-notch machines combined with advanced statistical mechanics techniques aimed at mimicking the Salvinia leaves and revealing its strategies for airtrapping. As will be shown, the results, obtained by exploiting the full potentialities of the Tier-0 computer architectures made available through the the WETMD project allocated by PRACE, have the potential to inspire next generation, biomimetic, superhydrophobic surfaces, as well as to provide benchmarks for continuum models of wetting and cavitation.

MOLECULAR SIMULATIONS

Chair: Ilpo Vattulainen, Department of Physics at the Tampere University of Technology, Finland

Title: Efficient Lennard-Jones Lattice Summation Techniques for Lipid Bilayers

Speaker

  • Erik Lindahl, Department of Biochemistry and Biophysics at the University Stockholm

Abstract
Surfaces exhibiting extraordinary features exist in nature. A remarkable example is the Salvinia molesta. This water fern, due the presence of small hydrophilic patches on top of rough hydrophobic surfaces, is able to retain air pockets when submerged by stabilizing the resulting Cassie state against positive pressure fluctuations while, at the same time, preventing bubble detachment. A similar strategy is adopted by certain insects and spiders (e.g. Elmis Maugei and Dolomedes triton) to breath underwater, thanks to a stabilized air layer, the so-called plastron. However, since CWT is a rare event beyond reach of brute force computations, the mechanism of wetting remains elusive and it is still difficult, if not impossible, to predict the transition from the Cassie to the fully wet Wenzel state. Using specialized techniques, it has been recently demonstrated that molecular dynamics is indeed capable to describe the Cassie-Wenzel transition on a simple model system. However, going beyond this proof-of-concept simulations, with the goal of reproducing real hydrophobic coatings and the complex morphology of natural surfaces, requires to combine a smart theoretical approach with a boost in computational resources. We discuss here the results of massively parallel simulations on top-notch machines combined with advanced statistical mechanics techniques aimed at mimicking the Salvinia leaves and revealing its strategies for airtrapping. As will be shown, the results, obtained by exploiting the full potentialities of the Tier-0 computer architectures made available through the the WETMD project allocated by PRACE, have the potential to inspire next generation, biomimetic, superhydrophobic surfaces, as well as to provide benchmarks for continuum models of wetting and cavitation.The introduction of particle-mesh Ewald (PME) lattice summation for electrostatics 20 years ago was a revolution for membrane simulations. It got rid of horrible cutoff effects, and removed the electrostatics cutoff’s influence on important properties such as area and volume per lipid. However, over the last decade it has become increasingly obvious that the Lennard-Jones cutoff is also highly problematic. Dispersion corrections are not sufficient for membranes that are neither isotropic nor homogenous – altering the cutoff will still alter properties. Here I will present a new highly efficient and parallel technique for LJPME that is part of GROMACS version 5. We have solved the historical problem with Lorentz-Berthelot combination rules in lattice summation by introducing a series of approximations, first by using geometric combination properties in reciprocal space, and now also correcting for this difference in direct space terms. Not only does this improve molecular simulation accuracy by almost an order of magnitude, but it also achieves absolute LJPME simulation performance that is an order of magnitude faster than alternatives – in many cases it is within 10% of the previous cutoff performance in GROMACS.

Title: On the Activation and Modulation of Voltage Gated Ion Channels

Speaker

  • Mounir Tarek, CNRS & Université de Lorraine

Abstract
Excitable cells produce electrochemical impulses mediated by the transport of ions across their membrane through proteins pores called ion channels. The most important family of channels propagating an electrical signal along the cell surface is the voltage-gated ion channel (VGCs) family. VGCs are essential physiological effectors: they control cellular excitability and epithelial transport. A myriad of genetic mutations found in the genes encoding their subunits cause channel malfunction. These so-called channelopathies have been incriminated in a variety of diseases, including, among others, epilepsy, pain syndromes, migraines, periodic paralyses, cardiac arrhythmias, hypertension and hypotension. Contemporary research will benefit from new insights into the minute molecular details in play which can contribute to a fine understanding of VGCs function, as well as its modulation by the environment or its disruption by specific mutations. The working cycle of VGCs involves the complex conformational change of modular protein units called voltage sensor domains (VSDs). For over forty years, these rearrangements have been recorded as “gating” currents, intensities and kinetics of which are unique signatures of VGC function. In this Talk we show that the atomistic description of VSD activation obtained by molecular dynamics simulations and free energy calculations is consistent with the phenomenological models adopted so far to account for the macroscopic observables measured by electrophysiology. Most importantly, by providing a connection between microscopic and macroscopic dynamics, our results pave the way for a deeper understanding of the molecular level factors affecting VSD activation, such as lipid composition, amino acid mutations, and binding of drug molecules or endogenous ligands.

Title: Effect of Hydrophobic Pollutants On The Lateral Organization Of Biological Model Membranes

Speaker

  • Luca Monticelli, Institut de Biologie et Chimie des Protéines CNRS

Abstract
Cell membranes have a complex lateral organization featuring domains with distinct composition, also known as rafts, which play an essential role in cellular processes such as signal transduction and protein trafficking. In vivo, perturbation of membrane domains (e.g., by drugs or lipophilic compounds) has major effects on the activity of raft-associated proteins and on signaling pathways. In live cells, membrane domains are difficult to characterize because of their small size and highly dynamic nature, so model membranes are often used to understand the driving forces of membrane lateral organization. Studies in model membranes have shown that some lipophilic compounds can alter membrane domains, but it is not clear which chemical and physical properties determine domain perturbation. The mechanisms of domain stabilization and destabilization are also unknown. Here we describe the effect of six simple hydrophobic compounds on the lateral organization of phase -separated model membranes consisting of saturated and unsaturated phospholipids and cholesterol. Using molecular simulations, we identify two groups of molecules with distinct behavior: aliphatic compounds promote lipid mixing by distributing at the interface between liquid-ordered and liquid-disordered domains; aromatic compounds, instead, stabilize phase separation by partitioning into liquid-disordered domains and excluding cholesterol from the disordered domains. We predict that relatively small concentrations of hydrophobic species can have a broad impact on domain stability in model systems, which suggests possible mechanisms of action for hydrophobic compounds in vivo.

HPC in Industry

This session took place Wednesday 27 May 2015.

HPC IN INDUSTRY IN IRELAND

Chair: Leo Clancy, Division Manager ITC, IAD Ireland

Title: Subsurface Imaging of the Earth for Exploration: Methods and HPC Needs

Speaker

  • Sean Delaney, Tullow Oil

Abstract
Imaging the subsurface of the Earth is a challenging task which is fundemental in determining resource location and management. Seismic wave propagation through the Earth is a key tool in geophysics and one of the best available methods for imaging the subsurface and studying physical processes in the Earth. Seismic imaging has moved from ray based imaging operators to numerical solutions to wave equations, usually termed full wavefield imaging. This methodology requires high performance computational (HPC) resources. The most recently implemented tools for imaging are Reverse Time Migration (RTM) and Full Wavefield Inversion (FWI). RTM imaging has been shown to be highly beneficial in imaging in complex geological regions whilst FWI has been used to develop high resolution velocity models, which leads to better subsurface images. The practical implementation needs expertise in imaging and in running HPC applications to reduce the overheads, not just in CPU costs but in turn-around times for quicker business decisions. As more expensive imaging tools are continually being developed, the size of the data being recorded in the field has approached the petabyte scale per survey. Thus, there is a continuing need for HPC, not just on cutting edge imaging algorithms but for updating traditional imaging codes. This presentation will discuss seismic imaging trends from a non-specialist point of view with focus on industry applications. The overall message being, the seismic exploration industry is still pushing the upper barrier of computational geophysics on HPC resources for processing and imaging methods. To extract the maximum benefit and throughput and minimise costs in CPU spend, a combination of cutting edge tools, advanced imaging specialists and geophysicists/physicists highly skilled in HPC are required. For more information, see www.tullowoil.com.

Title: The DNA Data DELUGE-A Parallel Computing Approach

Speaker

  • Bendan Lawlor, nSilico

Abstract
A Formula 1 engine is powerful, highly engineered and capable of tremendous numbers of revolutions per second. To win races with such an engine, it’s necessary to house it in a suitable chassis, and make sure that it gets fuel quickly. Similarly, a fast low-level algorithm, developed with a deep understanding of the target processor, needs to be correctly housed and fed in order to convert that sheer power into a useful solution. NSilico is an Irish based SME that develops software to the life sciences sector, providing bioinformatics and medical informatics systems to a range of clients. Processing the exponentially growing amount of genomic sequence data is a major challenge to those clients. The Formula 1 engine in question is a SIMD C-language implementation of the Smith-Waterman algorithm, presented as a project of the PRACE SHAPE programme to this conference last year. This talk outlines the technical challenges in harnessing this powerful implementation into a scalable, resilient service cluster at relatively low cost. The presented proof of concept solution uses the Scala language and the Akka framework to demonstrate how two primary abstractions – the Actor and the Stream – are suited to this task.

PRESENTATION PDF DOWNLOAD

Title: An Irish SME’s View of Big Data Analytics

Speaker

  • Dave Clarke, Asystec

Abstract
The presentation will focus on giving a perspective on the relevance of Big Data Analytics to an Irish SME. With various assessments of where big data is on Gartner’s Innovation Hypecycle, Dave’s talk will review some of the key development tracks that big data has taken over the last 5 years. He will note some of the key challenges that he sees for Irish SME’s notably the changing big data technology landscape and the need to collaborate with new people both internal and external to the organisation. These new stakeholders have diverse backgrounds and needs from DNA gene sequencing to supply chain administration to call centre service optimisation.

PRESENTATION PDF DOWNLOAD

HPC IN INDUSTRY

Chair: Lee Margetts, Research Computing Services The University of Manchester

Title: On the Impact of Automatic Parallelization in Technical Computing for Science and Industry

Speaker

  • Manuel Arenaz, University of Coruña & CEO of Appentra Solutions

Abstract
High Performance Computing is a key enabling technology to solve the big challenges of modern society and industry. The development of HPC programs is a complex, error-prone, tedious undertaking that requires a highly-skilled workforce well trained in HPC methods, techniques and tools. In the years to come, a large number of HPC experts are expected to retire. Thus, there is a growing urgency in mending the HPC talent gap, especially in market segments such as Oil&Gas and R+D/Government where HPC is competitive advantage. Parallelism is the primary source of performance gain in modern computing systems, and compiler technology is at the heart of many developer tools available in the HPC marketplace. Thus, automatic parallelization is a key approach to address the HPC talent gap as it decouples the development of HPC programs from the features and complexity of the underlying parallel hardware. Overall, parallelizing compilers enable experts to focus on computational science and engineering methods, techniques and tools, getting them rid of learning HPC methods, techniques and tools. Automatic parallelization is a long-lasting challenge for the HPC community. There have been many efforts world-wide in academia and industry to build parallelizing compilers that effectively convert sequential scientific programs into parallel-equivalents. Well-known examples are ICC, PGI, GCC, Polaris, SUIF, Pluto,… Recent advances in compiler technology to automatically extract parallelism from sequential scientific codes have solved the limitations of classical dependence analysis technology. A new hierarchical dependence analysis technology has been transferred from academia to industry by Appentra Solutions. The resulting product is Parallware, a new source-to-source parallelizing compiler for C programs that supports the OpenMP parallel programming standard. This talk will analyze the state-of-the-art in parallelizing compilers as well as their impact in modern science and industry from the point of view of performance, portability and productivity.

See the video presentation here.

Title: A Self-Managed HPC Cloud Ecosystem for Simulation and Collaboration

Speaker

  • Nicolas Tonello, Director, Constelcom

Abstract
Simulations, whether for virtual engineering, life sciences, or data processing and analysis in general are becoming essential to the creation of innovative products and scientific discoveries. Simulating ever larger and more complex problems to replace experiments or prototyping requires the kind of High Performance Computing (HPC) power traditionally only available in national research centres. However, whilst HPC capability is growing very fast, true supercomputing remains a specialist area often delegated to computing specialists as opposed to engineering and science discoverers, which restricts its uptake. In this talk, we will present some of the ideas and concepts which we have evolved over the last five years to address the challenges and common requirements for all end-users and simulation applications, in order to foster Highly Collaborative Computing (HCC) supported by HPC and encourage utilisation by a much wider community of non-specialists through ease of access, ease of use and self-management, and which have led to the deployment of the first instance of ConstellationTM on the Hartree Centre’s systems in the UK. The key elements and the need for close collaboration and interaction with supercomputing centres in order to develop this HPC Cloud ecosystem will be discussed, as well as plans for future developments and possibilities as wells as challenges to expand and create a pan-European connected community of users and resources.

Title: Establishing HPC Computational Environments in Industry: A View from the Inside

Speaker

  • Stefano Cozzini, CEO eXact Lab

Abstract
Stefano Cozzini will present the experience of an innovative startup to provide HPC computational environment in industry and beyond. eXact lab srl, founded just three years ago with the aim to provide High performance Computing services is still in its start-up phase but it is gaining experience and learning how to promote and establish what we define as an HPC computational environment f outside a purely research and academic world. Our idea of a computational environment will be discussed and some successful case studies illustrated. He will then close discussing the challenges ahead of us.

PRESENTATION PDF DOWNLOAD

Keynotes & Panels

This session took place Thursday 28 May 2015.

Title: International Technology Investment & HPC

Speaker

  • Leo Clancy, Division Manager ITC, IAD Ireland

Abstract
Leo Clancy will focus in his presentation on global trends in international technology investment. His work at IDA (Industrial Development Authority) Ireland means that he has significant experience of this area. He will outline some of the areas where he sees HPC becoming more and more important in terms of these trends.

PRESENTATION PDF DOWNLOAD

Title: Ultimate Rayleigh – Benard and Taylor-Couette Turbulence

Speaker

  • Detlef Lohse, Faculty of Science and Technology University of Twente

Abstract
Rayleigh-Benard flow – the flow in a box heated from below and cooled from above – and Taylor-Couette flow – the flow between two coaxial co- or counter-rotating cylinders – are the two paradigmatic systems in physics of fluids and many new concepts have been tested with them. They are mathematically well described, namely by the Navier-Stokes equations and the respective boundary conditions. While the low Reynolds number regime (i.e., weakly driven systems) has been very well explored in the ‘80s and ‘90s of the last century, in the fully turbulent regime major research activity only developed in the last decade. This was also possible thanks to the advancement of computational power and improved algorithms and nowadays numerical simulations of such systems can even be done in the so-called ultimate regime of turbulence, in which even the boundary layers become turbulent. In this talk we review this recent progress in our understanding of fully developed Rayleigh-Benard and Taylor-Couette turbulence, from the experimental, theoretical, and numerical point of view, focusing on the latter. We will explain the parameter dependences of the global transport properties of the flow and the local flow organisation, including velocity profiles and boundary layers, which are closely connected to the global properties. Next, we will discuss transitions between different (turbulent) flow states. This is joint work with many colleagues over the years, and I in particular would like to name Siegfried Grossmann, Roberto Verzicco, Richard Stevens, Erwin van der Poel, and Rodolfo Ostilla-Monico.

Title: Panel: Science and Industry – Partners for Innovation

Moderator

  • Tom Wilkie, Scientific Computing World

Panelistrs

  • Sylvie Joussaume, INSU/CNRS
  • Anders Rhod Gregersen, Vestas, and Vice-Chair of the PRACE Industrial Advisory Committee
  • Mateo Valero, Director Barcelona Supercomputing Center
  • Augusto Burgueño Arjona, European Commission

Abstract
During the last three days speakers from industry and academia presented results that would have not been achievable without HPC. Will continuing on the proven and successful track be sufficient to meet the goals that Europe sets itself to be a leader in innovation and scientific excellence. The panelists will discuss the respective expectations of industry, science, infrastructure providers, and funding agencies on future HPC technologies and services. They will explore opportunities for synergies and mutual transfer of know-how between the stakeholders. Questions from the audience are welcome and should be submitted to the panel chair prior to the session or via the open microphone during the discussion.

PRACEdays15 Highlights

Interview Lee Margetts

PRACE Industrial Advisory Committee PRACE RI

Interview with Jean Christoph Desplat

Director of ICHEC PRACE RI

Exascale Highlights

Satellite event during PRACEDays 15 – 26 May 2015 PRACE RI

Interview Leo Clancy

Head of Technology Consumer & Business Services AT PRACE RI

Interview Seki Mashiro

RIST Japan

Interview Panchatcharam Mariappan

Best Poster Award

PRACEdays15 Posters

The following Posters were presented at the PRACEdays15 Poster Sesssion.

Poster Title: Winner of the PRACEdays15 Best Poster Award: GPU accelerated finite element method for radio frequency ablated cancer treatment

Poster Authors

  • Panchatcharam Mariappan, NUMA Engineering Services Ltd, Ireland
  • Phil Weir, NUMA Engineering Services Ltd, Ireland
  • Ronan Flanagan, NUMA Engineering Services Ltd, Ireland

Abstract
Graphics Processing Units (GPUs) are nowadays used for numerical computation , beyond their original purpose of graphics accelerators. Mature hardware and GPU software tools and libraries support double precision and memory correction. GPU accelerated computational fluid dynamics has gained attention in both academia and industry. In this article, we investigate the importance of GPUs as accelerators in the field of biomedical engineering. We developed a software tool to predict the lesion development in cancer patients after the radio frequency ablation cancer treatment. We use Penne’s bioheat model with appropriate boundary conditions and the finite element method for numerical discretization. From the finite element discretization of the bioheat equation, we observe that no explicit element integration is required. Since the problem domain is fixed, we find the neighbours of each node at the first time step and generate a compressed sparse row structured (CSR) matrix which can be used for the entire domain. After the CSR matrix is generated, we send the domain information such as nodes, elements and matrix information (e.g. the CSR matrix rows and columns) to the GPU. The Central Processing Unit (CPU) loads the initial data, finds the neighbours list, generates the CSR matrix and stores the results on the disk, whereas the GPU constructs the shape functions, assembles the local stiffness matrix into the global matrix in the CSR form and solves the sparse linear system with the help of the existing CUDA libraries such as CUBLAS and CUSPARSE. In order to solve the linear system, we employed the ILU preconditioned BiCGStab algorithm, one of the fastest solvers among Krylov subspace solvers. At each time step, the GPU generates the heat source term and solves the cell death model, while the CPU saves the results in vtu/vtp files. The heat source term generation is based on our in-house point source model for approximating the Joule heating effect, and the cell death model is an adapted evolution equation, predicting whether cells near the tumour are alive or dead. The tasks assigned to the GPU are the most time consuming parts of the finite element method and the GPU accelerates them with the desired speed-up and accuracy. The major steps involved in this work are receiving the segmented CT scans of the patient from the doctors, generating the mesh, obtaining the needle position from the CT scans (approximately the centre of the tumour) and simulating them using our software tool. Existing software tools working on multi-core CPUs (Intel i5’s) take 6 hours to predict a lesion for 26 minutes of real treatment time, for around 1 million elements. Our current work with the assistance of the GPU acceleration yields the result in approximately 3 minutes for the same number of elements, where the comparison is done with Intel Xeon CPU E5 – 2680 @ 2.8 GHz, and NVIDIA GeForce Titan Black GPU @ 3.5 GHz (2880 CUDA Cores).

POSTER PDF DOWNLOAD

Poster Title: Numerical study of mixing in swirling coaxial jets – An application of large eddy simulation

Poster Authors

  • Teresa Parra Santos, Dept. of Energy and Fluid Mechanics, University of Valladolid, Spain/li>

Abstract
Assessment of Large Eddy Simulation (LES) models of confined coaxial swirling jets is the aim of this work. The mid-term application is to improve the stabilization of flames of poor mixtures by means of a swirling flow. This provides saving of fuel as well as a reduction of contaminant emissions. Swirling burners have some advantages when compared with bluff bodies and cross flows. These are lower head losses and soot, less maintenance tasks. A possible application of the smallest burners is the food industry since baking with low emissions improves the taste of the product.

Despite the simple geometrical set-up of the benchmark, the flow pattern shows complex aerodynamic behavior. The simple burner considers the use of two coaxial nozzles: one axial with fuel and another annular with air. The expansion of the flow, when entering the chamber will produce the Outer Recirculation Zone. If swirl number is large enough to let the flow turn back into the centre, the vortex breakdown phenomenon appears to form an Inner Recirculation Zone limited by two stagnation points located in the axis of the chamber. The region between both recirculation zones with high shear is where mixture of fuel-air occurs.

This work is devoted to isothermal flow to gain an insight of flow pattern associated to different swirl numbers and diffusers. Axial swirl injector is composed by 8 fixed vanes in the annular nozzle. It is responsible of the azimuthal momentum of annular jet. The Swirl number is associated to the angle of the trailing edge of the vanes. Besides, the influence of conical diffusers in the mixture is analyzed.

Swirl numbers of 0.2 (low) 0.6 (intermediate) and 1.2 (strong) were tested. To sum up, the strong swirl number had the lead stagnation point near the discharge of the nozzles and provided a mixing length lower than half diameter of the chamber. Intermediate swirl number have bigger Outer Recirculation Zones and the mixing length is more than one diameter. Finally the low swirl number does not have any vortex breakdown and the mixing length is several diameters.

Bearing in mind the influence of conical diffusers, it is more important in the case of intermediate swirl numbers since the diffuser reduces the mixing length. A 140º diffuser is able to avoid the Outer Recirculation Zone for Strong Swirls. The same diffuser setup operating with intermediate swirl number is able to prevent the formation of Taylor-Couette instabilities (counterrotating vortex rings) with the associated reduction of head losses.

These models were tested using the LES algorithm Scale Selective Discretization scheme. Temporal resolution is 10-5 s/timestep with spatial resolution 5 times larger than Kolmogorov scale. It was found that for mesh of 7-9 million cells without multigrid, the optimum is 64 processors. If the multigrid set up is modified to consider more cells in the coarsest level, the optimum number of processors can be increased to 128. Also, the increase of tolerance has an impact for efficient use of larger number of processors.

The author thankfully acknowledges the Spanish Ministry of Science and Innovation for the financial resources in the framework of the project reference ENE2011-25468. We acknowledge PRACE for awarding us access to resource Curie-GENCI@CEA based in France and MareNostrum@BSC based in Spain. Ref. 2010PA1766

POSTER PDF DOWNLOAD

Poster Title: Numerical simulation of non-premixed swirling flames

Poster Authors

  • Ruben Perez, Dept. of Energy Engineering and Fluid Mechanics, University of Valladolid, Spain
  • Teresa Parra, Dept. of Energy Engineering and Fluid Mechanics, University of Valladolid, Spain

Abstract
The present work focuses on the numerical simulation of diffusive flames in a confined swirl burner. The background motivation for the project arises from the greenhouse gas emissions. In methane operated burners, the methane slip due to incomplete combustion is a problem since methane is a harmful greenhouse gas. Lean flames produce less contaminant emissions and reduce fuel consume, however they are unstable. The swirling flow is a stabilizer of the flame so that poor mixtures can be burned.

The governing equations for 3D, transient, reactive flow are solved with a second order scheme. The 3D mesh has 4 million hexahedral cells. No multi-grid was used for reactive cases because the averaging on temperature field is a precursor of the lack of accuracy on the reaction rate. As for the turbulence model, the k-ε was selected.

Numerical model for no swirl and high swirl burners have been carried out using heat and mass transfer for non reactive cases and a simplified mechanism of reaction for the reactive case. Regarding the reactive case: three stoichiometries were mixed and burned, stoichiometric (lambda = 1), lean (lambda = 1.2) and rich (lambda = 0.8) mixtures. The temporal resolution must be around 10-7 s/timestep, because of the stiffness of the reactive case.

Contrasting non reactive and reactive cases, the last one produces higher axial velocities to keep the mass balance. Hence, it is a precursor of smaller Inner Recirculation Zones (IRZ) in the case of strong swirls. The lead stagnation point of the IRZ plays an important role fixing the location of the flame front in swirling burners. Besides, the hot products of reaction of the IRZ help to warm the fresh mixture. Contrasting flames with swirl number null, 0.6 and 1 it is possible to conclude the decrease of the flame front thickness while increasing the swirl number.
Contrasting different stoichiometries, lean mixtures have lower equilibrium temperature and therefore, the thermal emission of nitrogen oxides is lower. However, strong swirls are needed for very poor mixtures in order to be burn in a stable way.

The author thankfully acknowledges the Spanish Ministry of Science and Innovation for the financial resources in the framework of the project reference ENE2011-25468. We acknowledge PRACE for awarding us access to resource Curie-GENCI@CEA based in France and MareNostrum@BSC based in Spain. Ref. 2010PA1766

POSTER PDF DOWNLOAD

Poster Title: AMOEBA and HPC: Parallelising Polarisable Force Fields

Poster Authors

  • Weronika Filinger, EPCC, The University of Edinburgh

Abstract
For decades now classical and quantum mechanical based computational chemistry codes have been used to evaluate molecular properties and interactions in gas and condensed phases. However, there are a number of systems that cannot be simulated correctly using the currently available software as they are limited by the use of non-polarisable classical models. These models approximate electrostatic interactions by only using fixed charges, which means that the charges in atoms are not allowed to adapt to changes in their local environment. Hence these models are not capable of providing sufficiently accurate results for systems where the polarization and self-induced charges play a significant role in the interaction. Polarisation changes the geometry and energy of the system by distorting the bonds between atoms and it is a necessary intermolecular interaction when modelling chemical reactions in biological molecules and water-based systems. Therefore, to tackle a wider range of problems in chemistry, biochemistry and material science where charges and polarization are important it is necessary to implement more realistic modelling of molecular forces that take into account these effects.

Thus efforts have been put into developing mutually polarisable models that enhance the fixed charge models by including the polarisation of atoms and molecules in the dynamics of these systems. These polarisable models are more accurate and more complex hence more computationally demanding. Typically the size and number of time steps in these simulations are selected so that the calculation can finish within a reasonable time period but they also need to be long enough to be relevant to the timescales of the process being simulated. For example, most proteins and DNA simulations need to span at least nanoseconds of simulated system time, which can take days to years of wall clock time on a single CPU. Moreover, the use of polarisable fields, which are necessary for most biological systems, will increase this time by ten-folds or more. Hence, there is a strong need to develop the software packages implementing polarisable force fields that are capable of exploiting current and emerging HPC architectures to make modelling ever more complex systems viable.

One of the more prominent polarisable models in this field is AMOEBA (Atomic Multipole Optimised Energies for Bimolecular Applications). This has been implemented and is widely used in codes such as TINKER and AMBER. AMOEBA replaces fixed partial charge with atomic multipoles and includes an explicit dipole polarisation, which allows atoms to respond to the polarisation changes in their molecular environment. AMOEBA has been shown to give more accurate structural and thermodynamic properties of various simulated systems. However, the time needed to calculate the effect of the induced dipoles alone is 15-20 times larger than the cost of one time step for a fixed charge model. If the computational time could be reduced then the AMOEBA force field could, for example, model a much broader range of protein-ligand systems than is currently possible or provide a means of refining in X-ray crystallography for large and challenging data sets such as ribosome crystals.

The larger computational costs associated with the calculation of polarisations needs to be mitigated by improved the parallelisation of codes such as TINKER. In this poster we give more details on the AMOEBA model as implemented in the TINKER molecular modelling package and present our efforts at improving its parallelisation through the use of hybrid MPI and OpenMP techniques.

POSTER PDF DOWNLOAD

Poster Title: Novel multiphase simulations investigating cavitation by use of in-situ visualization and Euler/Lagrange coupling

Poster Authors

  • M. Bode, Institutefor Combustion Technology, RWTH Aachen University, Aachen, Germany
  • J.H. Göbbert, JülichAachen Research Alliance (JARA-HPC), RWTH Aachen University, Aachen, Germany
  • H. Pitsch, Institutefor Combustion Technology, RWTH Aachen University, Aachen, Germany

Abstract
Flow configurations involving both liquid and gaseous fluids often occur in industrial applications. Engine injection systems, which are used to atomize liquid fuels, are one example. The performance of such atomizers depends on a cascade of physical processes, originating from the nozzle internal flow, cavitation, turbulence, and the mixing of a coherent liquid stream with a gaseous ambient environment. The transfer occurring between liquid and gas is governed by an interface topology. An accurate prediction of this mixing stage is crucial for reliable predictions of the overall system as it is typically the first physical process to be modeled in simulations and uncertainties here will influence, for example, the design and performance of engines all the way down to emission and pollutant formation.

Within the last years, engine experiments have shown that the impact of cavitation on the overall engine processes is much larger than current knowledge would predict. Due to the small size of injection systems, which have outlet diameters on the order of 100 micrometers, and the resulting bubbles, droplets and turbulence structures, which are even much smaller, a detailed investigation using experiments is very difficult and simulations can be quite helpful.

Accurate simulations of the whole atomization process have to include a broad spectrum of different length scales and resolve the interface topology in an exact and robust way, which is hard to achieve even with massively parallel code frameworks on TIER-0 HPC systems. However, recent developments with respect to interface tracking methods and a new simulation approach combining Eulerian and Lagrangian spray simulation techniques in order to decrease the computational cost in physically less important flow regions, enables us to study the impact of cavitation on the mixing process. Additionally, new in-situ visualization techniques enable a smart data management, which stores fully resolved data only in important flow regions leading to a higher information/data size ratio, which is crucial for model development.

This work presents the new simulation techniques as well as its application to realistic atomizers. The CIAO code framework was used on MareNostrum III and JUQUEEN for data generation and the data are studied focusing on the effect of cavitation on commonly used spray models in industrial context.

POSTER PDF DOWNLOAD

Poster Title: Massively parallel code to search for gravitational waves from rotating neutron stars in advanced detector data

Poster Authors

  • G. Poghosyan, Steinbuch Centre for Computing of KIT
  • S. Matta, PEC University of Technology
  • A. Streit, Steinbuch Centre for Computing of KIT
  • J. Bolek, Faculty of Physics of WUT
  • M. Bejger, CAMK
  • A. Królak, IMPAN of Polish Academy of Sciences

Abstract
Gravitational waves are the last prediction of general relativity still awaiting a direct experimental verification. Observations of gravitational waves will open a new field – gravitational wave astronomy. First science data from the global network of advanced gravitational wave detectors – LIGO, GE0600 and Virgo long arm interferometers, are expected in July 2015. The advanced detector network will be sensitive to signals all over the sky, although source positions can be determined by triangulation. For these reasons, searching for sources in noisy data is algorithmically challenging, since one has to simultaneously look for different types of signals, and computationally formidable, due to the large parameter space over which the searches must be carried out.

To perform a rapid analysis of all data from the advanced LIGO and Virgo gravitational wave detectors’ network, hundreds of millions of CPU hours will be required — the code utilizing the potential of massively parallel supercomputers is therefore mandatory.

Polgraw-Virgo group in cooperation with the Simulation Laboratory for Elementary- and Astro-Particles have developed a highly scalable computation code parallel PollGrawAllSky, which enables the data analysis of all-sky searches for gravitational wave signals at large scales on acceptable time scales. Benchmarking of the code in framework of PRACE Preparatory access on a Cray XE6 system was performed to show efficiency of our parallelization concept and to demonstrate scaling up to 50 thousand cores in parallel. To estimate the computational requirements when current version of code is used for analysis, we have performed representative tests with the Gaussian noise data at different band frequencies. For example, a serial search for GWs in one particular detection day at only 4 frequencies 600, 1000, 1700 and 2000 will require a total of 20 thousand CPU hours of computation, which is more than two years on a single CPU and correspondingly the output generated by this simulation would be ca. 4 GB.

To face the big challenge of the analysis of all the data that will be collected from the advanced detectors expected to be available by the year 2018, we are developing a hybrid parallelized PollGrawAllSky code able to scaling much above current 50000+ cores. To enhance the scalability of execution of many computations in parallel, we combine many instances consisting of different PolGrawAllSky executions that use different numbers of parallel sub-tasks. This feature is implemented using the dynamic process creation and grouping framework of MPI, with different MPI sub-worlds also known as virtual groups that enables collective and dynamic communication operations across a subset of related tasks. The main PolGrawAllSky code with parallel sky loop is encapsulated into another code, named skyfarmer, equipped with internal scheduling and bookkeeping mechanism.

With further implementation of usage of coprocessors (hardware accelerators) like Graphical Processing Units in parallel code presently optimised to use only standard Central Processing Units we hope to reach scalability level allowing to analyse at least four times more resources, i.e., 1000 million CPU hours, to perform analysis of final data in 2018. These means performance of 1petaFLOPS computer working continuously for one year.

POSTER PDF DOWNLOAD

Poster Title: Towards a quantitative understanding of the quark–gluon plasma

Poster Authors

  • Jon-Ivar Skullerud, Department of Mathematical Physics, Maynooth University, Ireland
  • Gert Aarts, Department of Physics, Swansea University, UK
  • Chris Allton, Department of Physics, Swansea University, UK
  • Simon Hands, Department of Physics, Swansea University, UK
  • Maria-Paola Lombardo, INFN – Laboratori Nazionali di Frascati, Italy
  • Seyong Kim, Department of Physics, Sejong University, Korea
  • Sinead Ryan, School of Mathematics, Trinity College Dublin, Ireland

Abstract
At extremely high temperatures (100,000 times those in the core of the Sun), the strong interaction, which holds quarks and gluons together to form protons, neutrons and other hadrons, undergoes a dramatic change of character. The quarks and gluons are no longer bound together, but instead form a new phase of matter called the quark-gluon plasma. At the same time, the quarks that make up ordinary matter become effectively massless as the chiral symmetry of the quarks, which is broken in ordinary matter, is restored. This state of matter existed in the first second after the Big Bang and is currently being produced in collisions between heavy ions (gold or lead) at CERN and Brookhaven.

The FASTSUM collaboration has been carrying out large-scale Monte Carlo simulations of strongly interacting matter at temperatures both above and below the transition to the quark-gluon plasma. These simulations have employed anisotropic lattices, where the lattice spacing in the temporal direction is much smaller than in the spatial directions. This allows a good resolution for temporal correlators, which is crucial to obtaining results for transport properties and survival of bound states in the plasma.

We will show results obtained for the electrical conductivity and charge diffusion as well as for states consisting of heavy (charm and beauty) quarks: the melting temperatures of the latter may be used as a “thermometer” of the quark-gluon plasma. We also show results for nucleons demonstrating the effects of chiral symmetry restoration. These results have been obtained from our “second generation” data ensembles generated with the use of PRACE resources. We are currently in the process of generating “third generation” ensembles which will double our temporal resolution and provide the first step towards a continuum extrapolation of our results.

POSTER PDF DOWNLOAD

Poster Title: Edge-elements for geophysical electromagnetic problems – A new implementation challenge

Poster Authors

  • Octavio Castillo Reyes, Barcelona Supercomputing Center (BSC), Computer Applications in Science & Engineering – CASE – Geosciences Applications, Spain
  • Josep de la Puente, Barcelona Supercomputing Center, Computer Applications in Science & Engineering – CASE – Geosciences Applications
  • Vladimir Puzyrev, Barcelona Supercomputing Center, Computer Applications in Science & Engineering – CASE – Geosciences Applications
  • José María Cela Espín, Barcelona Supercomputing Center, Computer Applications in Science & Engineering – CASE – Geosciences Applications

Abstract
Electromagnetic Methods (EM) are an established tool in geophysics, finding application in many areas such as hydrocarbon and mineral exploration, reservoir monitoring, CO2 storage characterization, geothermal reservoir imaging and many others. The last decade has been a period of rapid growth of marine electromagnetics, mostly because of its industrial adoption.

The marine controlled-source electromagnetic (CSEM) method has become an important technique for reducing ambiguities in data interpretation in the offshore environment and a commonplace in the industry. In the traditional configuration, the sub-seafloor structure is explored by emitting low-frequency signals from a high-powered electric dipole source towed close to the seafloor. By studying the received signal, the subsurface structures could be detected at scales of a few tens of meters to depths of several kilometers.

On the other hand, in the Finite Element Method for solving electromagnetic field problems, the use of Edge-based elements (Nédélec elements) has become very popular. In fact, Nédélec elements are often said to be a cure to many difficulties that are encountered (particularly eliminating spurious solutions) and are claimed to yield accurate results. However, the state of the art is marked by a relative scarcity in practice of robust codes to simulate geophysical electromagnetic problems. It’s may be attributed to their theoretical and implementational threshold. Indeed, more care and effort are required to implement them: basis functions, Piola mapping, edge directions and numbering strategy. Latter issues poses additional challenges.

Furthermore, the resultant data volumes of large-scale 3D modeling and simulations can easily overwhelm single core and modest multi core architectures. As a result, this kind of problems requires massively parallel computational resources in order to achieve a time frame acceptable to exploration process.

Based on previous ideas and considering the societal value of exploration geophysics, since this process is essential to among others, we present a novelty parallel implementation of Nédélec Elements for geophysical electromagnetic problems on unstructured meshes in 2D and 3D. The usage of unstructured meshes and mesh refinement make it possible to represent complex geological structures precisely and to improve the solution’s accuracy.

In particular, we present a simple, flexible and parallel implementation for Edge Elements in anisotropic mediums. The described software stack relies on a flexible solution which allows a general point of view. The efficiency and accuracy of the code is evaluated through a convergence test, scalability test, assembly time, and solver time, with a strong emphasis on the performance when the number of elements and degrees of freedom grows.

Since our target application is exploration geophysics, the results of this research stage shapes the future line of work to solve more complex problems such as forward modeling simulations and domain and functional decomposition.

POSTER PDF DOWNLOAD

Poster Title: Parallel agent-based simulation of South Korean population dynamics

Poster Authors

  • Josep Casanovas, Barcelona Supercomputing Center Universitat Politècnica de Catalunya
  • Cristina Montañola Sales, Barcelona Supercomputing Center Universitat Politècnica de Catalunya
  • Dr. Chang-Won Ahn, South Korea Electronics and Telecommunications Research Institute

Abstract
Changes in our society have created a challenge for policymakers, who confront a need of tools to evaluate the possible effects of their policies. Agent-based simulation is a promising methodology that can be used in the study of population dynamics. However, it has been little used in demographic research to help explaining dynamics. Simulation methodologies provide the opportunity to develop a virtual laboratory for exploring and validating current and new approaches. The purpose is to avoid conducting real social experiments, which may be expensive, unethical or even infeasible.

Agent-based simulation is commonly used for small scenarios because the number of agents and interactions between them can be extremely large in some of case studies, thus forcing the scientist to limit its number in order to execute the simulation in a standard computer. However, in the case of policy models, both the amount of compute power required and detailed micro-level data are significant. To deal with complex social models we can take advantage of parallel computation. Traditionally, parallel simulation has been applied in numerous scientific simulations such as networks or military. Nevertheless, the number of applications in the social sciences is scarce. One of the obstacles hindering the use of agent-based simulation is its scalability, especially if the analysis requires large-scale models. Currently there is no consensus on how to compute agent-based simulations in High Performance systems. Scalability issues cannot be solved just by distributing the computer workload on High Performance architecture. It depends on many factors, notably the execution platform and the complexity of the agent-based model, which in turn depends on the number of agents, their intelligent behavior and the complexity of their communication. It is not only important to address size problems but also to see whether more realistic agent-based models with complex behavior and communication network can be scaled-up to provide empirically and practically useful results.

A possible solution for scalability issues is to run the agent-based models on top of a scalable parallel discrete-event simulation engine. In our work, we use this solution in the design and development of a simulation framework that gives support for modeling and simulating agent-based demographic systems. It provides the placeholders for different demographic processes such as fertility, mortality, change in economic status, change in marital status, and migration. The main advantage of this approach is the ability to run large agent-based scenarios in High Performance Computing environments, when other current tools present limitations.

Moreover, we present a case study on forecasting demographics of South Korea during 100 years. South Korea is a country that shows the most unprecedented speed of aging in history. According to the latest projections, by 2050 South Korea may be the oldest country on earth. This situation could bring difficult challenges to face for the Korean government. Unless the country takes adequate measures to prepare for the demographic aging trend, it is expected that Korea will face a slower economic growth and living standards stagnation. With the application of agent-based simulation to this case we show how the life course of individuals is evolving, allowing deepen on the movements, interactions, and behaviours of South Korean population. Our model is able to capture individual characteristics and to overcome some data-related limitations with assumptions on behavioural rules. With this real case scenario, we show the potential of parallel agent-based simulation methodology for demographics.

POSTER PDF DOWNLOAD

Poster Title: Massively-parallel molecular simulation studies of ice and clathrate-hydrate nano-crystal and pre-cursor formation

Poster Authors

  • Niall English, Univ. College Dublin, Ireland

Abstract
Ice growth and decomposition was studied upon approximately spherical ice nano-particles of varying size surrounded by liquid water and at a variety of temperatures and pressures. The system box size was also varied for systems containing of the order of one million water molecules to almost ten million molecules, in order to establish system-size effects upon the growth and dissociation kinetics. It was found that there was a dependence upon system size on growth and dissociation, which points out the limitations of previous earlier simulation attempts in smaller simulation boxes.

Further, supercooled liquid water simulations were performed at various system sizes of the order to one to ten million water molecules, and the subtle re-arrangement of the structure and local density was explored as the system began to transition towards local ice-like conditions. Crucially, a system-size dependence was found upon these structural and dynamical rearrangements, which has been neglected in previous simulations.

The heterogeneous nucleation of ice nano-scale particles and the homogeneous nucleation of methane clathrate hydrates at water-methane interfaces were studied, again addressing the key question of the effect of system-size upon the results. It was found that both phenomena did depend on system-size, and that the subtle interplay between the frequency of box fluctuations and dilations with the underlying molecular rearrangements towards free-energy basins was quite important on influencing the outcome.

In the future, I would hope to continue these studies of clathrate and ice nucleation, growth and dissociation, especially with a view towards engineering applications, like the use of inhibitor compounds and temperature-/pressure-pulse strategies to regulate kinetics. Large-scale supercomputing is required to study these complex non-equilibrium processes without being plagued by the tyranny of small systems and periodic boundary conditions affecting results adversely. I expect that benefits to society will emerge from the greater understanding of these phenomena on a microscopic level, and the greater possibilities of devising kinetics-regulation strategies, e.g., to avoid pipeline blockage by hydrate plugs by inexpensive initial screening on supercomputing platforms, using molecular dynamics as an initial ‘predictive’ design tool.

POSTER PDF DOWNLOAD

Poster Title: Use of Graphics Cards (GPU) to Simulate Atoms, Molecules and Nucleus

Poster Authors

  • José Manuel Alcaraz-Pelegrina, University of Córdoba (Universidad de Córdoba), Department of Physics, Spain
  • Antonio Jesús Sarsa-Rubio, University of Córdoba (Universidad de Córdoba), Department of Physics, Spain

Abstract
The realistic description of the physical properties of the microscopic systems with a finite number of particles, such as nuclei or atoms and molecules, isolated or confined inside of molecular complexes, is a basic goal in Physics. These studies, even in the most basic aspects, lead to the use of complex methods that require powerful techniques of calculation being the Quantum Monte Carlo (QMC), one of the most developed in the last years. However, it is well known that QMC methods are computationally expensive.

Recently, a programming approach for performing scientific calculations on a graphics processing units (GPUs) has been developed. GPUs have evolved into a highly efficient data-parallel computing device and first market companies have released programming tools to make use of these technologies.

The implementation of QMC codes in GPUs will result in a better knowlegde of the microscopic systems such as nuclei or atoms and molecules.

A comparison between a GPU implementation of some QMC methods, a serial and a parallel code on CPU is presented for some systems.

POSTER PDF DOWNLOAD

Poster Title: Developing a scalable and flexible high-resolution code for direct numerical simulation of two-phase flows

Poster Authors

  • Lennon Ó Náraigh, University College Dublin, School of Mathematical Sciences, Ireland
  • Dr Prashant Valluri, Institute for Materials and Processes, School of Engineering, University of Edinburgh
  • Iain Bethune, Edinburgh Parallel Computing Centre, University of Edinburgh
  • Dr Toni Collis, Edinburgh Parallel Computing Centre, University of Edinburgh
  • David Scott, Edinburgh Parallel Computing Centre, University of Edinburgh
  • Mike Jackson, Software Sustainability Institute

Abstract
TPLS (Two-Phase Level Set) is an open-source program for simulation of two- phase flows in 3D channel geometries using high resolution DNS. TPLS solves the incompressible Navier—Stokes equations for a two-phase flow. A regular grid finite-volume discretization is employed based on an idealized channel geometry with a range of different inlet conditions that can be prescribed by the user. The interface between phases is tracked with a Level-Set method. The code evolves the physical variables (pressure, fluid velocities, and interface configuration) through discrete time steps. At each timestep, the key computational tasks performed amount to the solution of large systems of sparse linear equations with tens of millions of unknowns, for the key physical variables. In addition, regular I/O is required to save the system state for later analysis and visualizsation, or to restart in the case of hardware failure.

The code is implemented in Fortran90, initially with MPI parallelization using a 2D domain decomposition and bespoke Jacobi/SOR iterative solvers. Over the last two years, we have improved the TPLS code in several respects to give better performance, scalability and usability, moving from an in-house code specialized for use by the original developers, to an open-source flexible program which can easily be used by others, including academic and industrial users. The culmination of this work is TPLS version 2.0 (presented herein), where we have re-implemented the two most computationally-expensive solvers – the pressure and momentum steps – with calls to the PETSc library. Initial tests using the GMRES with Block-Jacobi preconditioner showed a speedup of 80% in the pressure solver on 2048 cores, along with improved strong scaling behavior. The original gather-to-master I/O strategy which wrote text files has been replaced with the use of NetCDF. As a result, we have obtained an order-of-magnitude reduction in I/O time, a compression factor of 6.7 and removed the memory bottleneck of requiring rank 0 to gather the entire domain. In addition to the Level Set method, we have added a PETSc implementation of the Diffuse Interface method (DIM), which is available as an option to users. Finally, with the support of the Software Sustainability Institute, we have added the ability to configure the code through input files or command-line arguments, obviating the need for users to modify and recompile for every application.

Novelty and Originality of Project: TPLS is unique in several aspects. Unlike other solvers (e.g. those mentioned below), TPLS solver has been purpose-built for supercomputing architectures like ARCHER. Most open-source solvers like OpenFOAM, Gerris, Fluidity and commercial solvers like ANSYS-Fluent/CFX offer only one interface-capturing method (the volume-of-fluid method) thereby limiting the applicability of these solvers to either free-surface, stratified, or wavy-stratified flows. The TPLS solver offers the users a choice of two types of interface capturing mechanisms between Diffuse-Interface Method and the Level-Set method. This enables the solver to accurately simulate a wide variety of physical scenarios. A further key feature of the work is the interdisciplinary composition of the development team, including researchers in HPC applications development, applied mathematics, algorithms design, and the physics of fluid flows.

The need for HPC in the work: Due to the high computational cost of a typical three-dimensional simulation, parallelization is essential, and scaling of the methodology described above has been demonstrated to several thousand CPU cores

Looking forward: TPLS has already been used to gain key fundamental insight into interfacial waves in two-phase flows, as well as in the hydrodynamics of evaporating droplets. Imminent future work will see TPLS applied to simulations of a much wider range of physical phenomena, with a view to gaining fundamental understanding of stratified-slug flow transitions, interfacial turbulence, contact-line motion, phase change, and heat transfer. We hope to enlarge the code’s user base, not only among academics, but also with industrial partners, including but not limited to the oil-and-gas industries.

POSTER PDF DOWNLOAD

Poster Title: Intelligent Water Drops Algorithm with Perturbation Operators for Atomic Cluster Optimization

Poster Authors

  • Ritchie Mae Gamot, University of Warwick, Centre for Scientific Computing, UK

Abstract
We present a modified version of the Intelligent Water Drops algorithm (MIWD) that has been adapted to allow it to be applied, for the first time, to global optimization of atomic clusters. Cluster perturbation operators were applied to further generate lower energies. The algorithm, dubbed as MIWD+PerturbOp, is an unbiased type of algorithm where no a priori cluster geometry information and construction were used during initialization which is not the case with other common search methods.

Four modifications were implemented: (a) The probability used to determine components for each agent in the population has factored in the pairwise potential energy; (b) The heuristic undesirability factor was based on an objective function used in a multi-start strategy study of LJ clusters by Locatelli, et al [2]; (c) The total worst agent in each iteration has also been identified, aside from total best agent, and paths belonging to it updated. (d) L-BFGS [3] was utilized to further relax clusters to its nearby local minimum.

Due to the iterative nature of the algorithm and the numerous combinations of parameters involved, HPC architecture was valuable in gathering results efficiently. Test runs reveal that a spherical bounding volume for the initial atom positions and grow-etch perturbation operator is a good combination for implementing final runs. Results achieved high success rates for sizes up to N = 104. This study outperformed the seeded runs of Basin Hopping with Occasional Jumping [4] in terms of success rates for more problematic clusters namely, LJ38, LJ75 – 77, LJ98, LJ101, and LJ103-104.

A detailed property analysis of the clusters of up to 104 atoms against the results in Cambridge Cluster Database (CCD) will be discussed. Preliminary experiments also show the method’s promise to binary LJ mixtures and Morse clusters which could be treated as a good indication of the method’s applicability to ionic, nanoalloy clusters or nanoparticles in general. Initial experiments on small Janus clusters using a potential model with a modified orientation term suited for two-patch Janus particles show promising configurations.

POSTER PDF DOWNLOAD

Poster Title: TemPortable task-based programming for Seismic Imaging

Poster Authors

  • Lionel Boillot, Inria, Magique3D team project, France

Abstract
Seismic imaging is of high interest for oil or gas detection. The most accurate techniques used by oil companies are the RTM (Reverse Time Migration) and the FWI (Full Wave Inversion). These methods are based on seismic wave simulations, ideally in anisotropic elastic media, a generally accepted realistic modeling of the subsurface. The parallelization of these methods is a laborious task. The main difficulty comes from the heterogeneity, at several levels. First, the mesh is generally unstructured and the mesh cells are non-homogeneous in terms of number of degrees of freedom. Second, the anisotropy and the boundary conditions leads to unbalance computational zones. These two points lead to heterogeneous zones with different number of computations in each zone. Finally, the hardware heterogeneity prevents to obtain balanced subdomains in a domain decomposition context because even knowing the exact number of computations does not imply to know the exact time they require to execute, due to the vectorization capabilities, the different memory caches,… In addition, the dependencies between the different computational subdomains create complex data movement patterns.

Current algorithms in shared memory use OpenMP to exploit many-cores architectures, with positive outcomes. Integration with CUDA (or OpenCL) allows for access to the computational power of accelerators. Going outside the node boundary, existing algorithms are using a message passing library to exchange data between nodes, enabling distributed memory versions. At this point, mixing these different approaches in order to achieve high performance across heterogeneous platforms remains a complex, error-prone and time-consuming task. The upcoming heterogeneous manycore revolution motivated by the race toward Exascale will only emphasize this problem. Other programming paradigms, especially task-based approaches, seem to be a suitable approach for such levels of hardware complexity. Task-based programming has been successfully applied for to many computational domains leading to robust and efficient solvers: in dense linear algebra (e.g. the DPLASMA library); in sparse direct and iterative methods and fast algorithms (e.g. FMM or H-Matrix problems). However, this programming paradigm has yet to be applied to large industrial parallel codes and demonstrated that these codes may be turned into portable and efficient task-based programs.

Our study focuses on the integration of the 3D elastodynamics Total code DIVA (Depth Imaging Velocity Analysis) with the PaRSEC runtime system, a specialized runtime for task scheduling, developed at University of Tennessee. The code was initially based on the MPI library for the parallelism. The principle of task programming relies on rewriting the code flow as task executions, each task being a computational function. The description of the tasks with the data they need in input and output forms a DAG (Direct Acyclic Graph) of tasks. The arrows of the DAG contain the data movement information that the runtime uses to determine the data flow exhibiting the parallelism.

We first addressed shared memory architectures with a ccNUMA (cache coherent Non Uniform Memory Access) node and an Intel Xeon Phi accelerator, based on the Intel MIC (Many Integrated Cores) architecture. The preliminary results are very promising since we obtain a positive speed up in comparison with the MPI-based code. The most interesting results concerns the parallel efficiency which decreased with the MPI-based code and which is stable and very close to one with the task-based code. In addition, the performance is portable on these two architectures. These results encouraged us to continue our work and move across the node boundaries into the distributed memory architectures and especially clusters of hybrid multicore nodes.

POSTER PDF DOWNLOAD

Poster Title: Dynamics of basic hydrolysis of methyl formate – Effect of microhydration

Poster Authors

  • Ivan Cernusak, Comenius University in Bratislava, Faculty of Natural Sciences, Department of Physical and Theoretical Chemistry, Slovakia

Abstract
Base catalyzed hydrolysis of methyl formate including one, two and three water molecules has been investigated using the ab initio molecular dynamics (MD) with the CP2K code. In MD calculations, we applied exchange functional by Perdew, Burke and Ernzerhof (PBE) with additional empirical dispersion correction (D3 method) by Grimme et al. For all atoms, basis sets of double-zeta quality with polarization in combination with the pseudopotentials proposed by Goedecker, Teter and Hutter (GTH) were used. The MD simulations were conducted at constant volume and temperature 298 and 350 K (NVT ensemble) maintained by CSVR thermostat with a time constant of 0.05 ps in the box of size 24 Å (open boundary conditions).

The study revealed that the important part of this micro-hydration assisted mechanism – the OH- attack on carbon in COH-group – strongly depends on the hydrogen-bonded network in the initial cluster. Several trajectories with different initial geometries of the hydrated CH3 O COH…OH- cluster and total length between 20-30 ps were analyzed. Only a fraction of these trajectories lead to traditional mechanism via BAC2 intermediate, while many end-up in stable but non-reactive hydrated ion coordinated to methyl formate.

Reaction profiles, geometry analysis and detailed participation of the solvent molecules (including animation of important trajectories) in selected cases are discussed.

POSTER PDF DOWNLOAD

Poster Title: Gender Inequality in HPC

Poster Authors

  • Athina Frantzana, The University of Edinburgh, EPCC, School of Physics and Astronomy, UK

Abstract
Gender inequality is a key problem across all scientific disciplines, both in academia and industry. HPC is a discipline that spans multiple traditional science subjects and relies on leading-edge scientific research. It would be plausible that the gender inequality issue, which has been identified and quantified across many fields of science and scientific research, is similarly present in HPC. To motivate action, we must measure and publicise the magnitude of the problem we face. To be effective, we must understand why women do not pursue careers in HPC so that our efforts can be appropriately targeted.

In 2014 the Software Sustainability Institute [1] produced a study on software development in Russell Group Universities [2]. This poster presents evidence of gender inequity in software development from further analysis of this study providing an initial insight into the group of people who use and develop HPC software. Our analysis shows that male researchers are more likely to develop software than female researchers (70% to 30%) and that men are more likely to receive training than women (63% to 39%). This has a profound effect on the people that developed their own software: 56% of the male respondents to the study had received training and developed their own software, whereas only 23% of the female respondents were trained and were developing software. Particularly important, as software developers in academia and HPC users are more likely to work with Linux or Mac. 66.2% of female respondents use Windows, with 33.8% using Mac, Linux or other operating system (Mac 25.7% , Linux 6.6 %). For the male respondents 42.2% use Windows, with the proportion of the male respondents using Mac (27.4%) or Linux (29.6%) doubling compared to men at 57.9%. The higher prevalence of Linux use in men and of Windows in women may be either the cause of the result of the lack of uptake in training by women. In this study we have also identified that women who responded to the survey are more likely to have less mature careers than men. Our analysis shows that there is not a considerable difference between the respondents that have less than 10 years of experience (66.9% of female and 57% of male). However, it is of great interest that 21.7% of the male respondents and only 9.7% of female have more than 20 years of experience.
This poster will present further details of this work and the potential impact of the findings on the HPC community.

[1] Software Sustainability Institute, http://www.software.ac.uk/
[2] Software Sustainability Institute (SSI) Russell Group study on software development, http://www.software.ac.uk/blog/2014-12-04-its-impossible-conduct-research-without-software-say-7-out-10-uk-researchers

POSTER PDF DOWNLOAD

Poster Title: The numerical Earth Magnetosphere

Poster Authors

  • Giovanni Lapenta, KU Leuven, Wiskunde, Belgium

Abstract
Coupling microphysics and macrophysics is the grandest of the grand challenges in almost all areas of science and engineering. Large scale system-wide effects can be triggered by the smallest scales in the system. An especially convenient field where the micro-macro coupling can be explored is the Earth space environment and its interaction with the incoming solar wind and cosmic rays. In this case we can access directly and measure both the system-wide scales as well as the smallest scales of interest. There are several space missions covering large regions of space surrounding the Earth (about 100 Earth radii, RE, in radius) capable of measuring the evolution of this environment. But from March 12, 2015 (the launch of the four spacecraft will be in about 12 hours at the moment of writing) we will have the Magnetospheric MultiScale (MMS) mission to probe the smallest scale of interest: the electron response scale that is about 100m in size. Needless to say simulating a sphere of 100 RE (or about 600,000 km) with a resolution of 100 meters is a grand challenge. A grand challenge that we at KU Leuven are attempting to solve using the best computing resources available via PRACE Tier-0 allocations.

In our recent work, we have consumed approximately 30 million CPU hours to achieve substantial advances that cover all major steps in the series of key processes developing in the Earth space environment. Like a bullet in the air, the Earth’s space cocoon (called magnetosphere) cuts through the solar wind flow, creating a shock wave that compresses and heats up the solar plasma. We are performing simulations of this interaction, never done before, successfully capturing the physics of the entire planetary environment in a domain of 160 x 120 x 120 planet radii, and capturing the detailed physics of individual electrons. Each global simulation requires 750 000 cpu hours, and is used to detect the variations of the particle velocity distributions and pressure anisotropies across the shock, and the effects of the local small scale particle physics on the large scale dynamics.

Within this domain we zoom in on the most important regions of energy release where magnetic reconnection develops. This process allows magnetic field to dissipate enormous energies that are suddenly released in bursts, a phenomenon that has defied explanation for decades. We have performed highly resolved simulations of the most important regions of interest in 3D using up to 48,000 cores and showing several new processes not identified before: the presence of switch-off shocks and rotational discontinuities, multiple interconnected sites each forming interacting islands of magnetic field and magnetic flux ropes embedding points of zero magnetic field (called magnetic nulls). Results that are published in the most prestigious journals in the field.

Every new Tier-0 simulation allows us to get closer to the real dynamics and scales of the plasma environment of our planet, which help us to better understand the impact of the Sun on our life and our technology. Our goal is to perform in the near future real scale, fully self-consistent, simulations of the full 3D environment of the Earth: the numerical magnetosphere.

Work supported by the European Commission via the DEEP project (deep-project.eu).

POSTER PDF DOWNLOAD

Satellite Events

The following three events took place during PRACEdays15:

Women in HPC

A Hands-on Introduction to HPC

This workshop took place Monday 25 May to Tuesday 26 May.

Instructors: Toni Collis, EPCC and Weronika Fillinger, EPCC

In collaboration with the Women in HPC network, the PRACE Advanced Training Centres (PATC) and the UK National Supercomputing Facility, ARCHER, the 1.5 day ‘Hands on Introduction to HPC’ training session took place. This course provided a general introduction to High Performance Computing (HPC) using the UK national HPC service, ARCHER, as the platform for exercises.

Familiarity with desktop computers was presumed but no programming or HPC experience was required. Programmers however gained extra benefit from the course as source code for all the practicals that were provided.

With all training staff being women, it provided an opportunity for women to network and build collaborations as well as learning new skills for a challenging and rewarding career in HPC.

Exascale Workshop - Enabling Exascale in Europe

Six European Exascale Projects – CRESTA, DEEP / DEEPER, EPiGRAM, EXA2CT, NUMEXAS and MontBlanc – run a compelling conference programme titled “Enabling Exascale in Europe for Industry” on Tuesday 26 May, 2015.

In six sessions, the projects demonstrated how their research efforts may benefit HPC usage in European industry. The speakers shared insights on their research efforts having in mind particularly user requirements and challenges to be overcome for future Exascale computing. They referenced concrete examples from various fields of industry, as for instance oil & gas, pharma or aerospace.

Additionally, the event included a keynote from a renowned external speaker, who shared first hand experience as a long-term HPC industrial user and shared his exascale vision: Dr. Eric Chaput, Senior Manager Methods & Tools Flight Physics AIRBUS.

The event especially focused on HPC users with a more industrial background, enlightening them on what will be possible in HPC in the future. Naturally, everyone interested in Europe’s road to Exascale computing benefited from this well thought-through programme was more than welcome to join, learn and discuss with us!

The Panel Discussion was led by Gilad Shainer, HPC Advisory Council Chairman.

Full details of the workshop can be found here.

Below, one can find the presentations of the workshop.

Title: KEYNOTE: Exascale Needs & Challenges for Aeronautics Industry

Speaker

  • Eric Chaput, Airbus, Flight-Physics Capability Strategy

Abstract
Exascale computing is seen as a key enabling technology for future aircraft design to be developed and optimised in a fully multidisciplinary way, making a wide use of design systems that provide integrated analysis and optimisation capabilities which allow for a realtime/interactive way of working. The move from RANS to unsteady Navier-Stokes simulation, (ranging from current RANS-LES to full LES) and/or Lattice Boltzmann method will significantly improve predictions of complex flow phenomena around full aircraft configurations with advanced physical modelling. For instance moving LES capability from Petascale to Exascale computing will accelerate the understanding of noise generation mechanisms and will enable the elaboration of flow control strategy for noise reduction. Multi-disciplinary analysis and design, and real time simulation of aircraft manoeuver, supported by affordable CFD-based aerodynamic and aero elastic data prediction will be a significant change of paradigm in aeronautics industry. The challenges faced by our industry at the horizon of 2025 will be presented together with the expectations on Exascale computing likely to bring operational benefits at that time.

PRESENTATION PDF DOWNLOAD

Title: DEEP & DEEP-ER: Innovative Exascale Architectures in the Light of User Requirements

Speaker

  • Estela Suarez, Jülich Supercomputing Centre; Marc Tchiboukdjian, CGG; Gabriel Staffelbach, CERFACS

Abstract
When developing new architectures for the Exascale era, the chicken-or-egg question arises of what to work on first: new hardware or new codes actually able to fully exploit Exascale systems. In the DEEP and DEEP-ER projects we tackle this challenge by adopting a comprehensive, holistic approach. We have come up with an innovative hardware concept, called the Cluster-Booster architecture. At the same time we develop the software stack and work very closely with our application partners to thoroughly integrate all three aspects. For our pilot applications, on the one hand we optimise their codes for our system, and on the other hand we are developing the system design based on the Exascale requirements that our users have. In this session we will explain our basic concept and share two of our use cases: Our industry partner CGG will talk about seismic imaging in the oil and gas industry, and our partner CERFACS on computational fluid dynamics. These two use cases will clearly demonstrate the potential the DEEP architecture offers at Exascale, not least for industrial users.

PRESENTATION PDF DOWNLOAD

Title: Mont-Blanc: High Performance Computing from Commodity Embedded Technology

Speaker

  • Filippo Mantovani, Barcelona Supercomputing Center

Abstract
In this session, the coordinator of the Mont-Blanc project will present an overview and status of this European project together with RR HPC Tech Lead Specialist-Aerothermal Methods at Rolls-Royce, a member of the Industrial End-User Group. He will present their observations from the process of testing the low-energy HPC prototypes produced by the project.

PRESENTATION PDF DOWNLOAD

Title: CRESTA: Developing Software and Applications for Exascale Systems

Speaker

  • Mark Parsons, EPCC, the University of Edinburgh

Abstract
The CRESTA project was one of three complementary Exascale software projects funded by the European Commission. The recently completed project employed a novel approach to Exascale system co-design, which focused on the use of a small set of representative applications to inform and guide software and systemware developments. The methodology was designed to identify where problem areas exist in applications and to use that knowledge to consider different solutions to those problems, which inform software and hardware, advances. Using this approach, CRESTA has delivered on all of its outputs, producing a set of Exascale focused systemware and applications.

PRESENTATION PDF DOWNLOAD

Title: EPiGRAM: Software in Support of Current and Future Space Missions

Speaker

  • Stefano Markidis, KTH Royal Institute of Technology

Abstract
During the preparation of NASA and ESA space missions, several simulations of different scenarios in space are carried out on HPC systems. These large scale simulations allow scientists to plan the space missions and to investigate possible phenomena of interest. In this talk, we present the new software developed by the EPiGRAM project to increase the scalability of these codes, the performance of the I/O activities and the amount of useful data for analysis. The impact of the EPiGRAM software on the current NASA Magnetospheric Multiscale Mission (MMS) and on the proposed ESA THOR mission (http://thor.irfu.se/) is discussed.

PRESENTATION PDF DOWNLOAD

Title: NUMEXAS: Embedded Methods for Industrial CFD Applications

Speaker

  • Riccardo Rossi, CIMNE – International Centre for Numerical Methods in Engineering

Abstract
A problem of paramount importance in the simulations of real engineering problems is the construction of a suitable discretization. It is widely acknowledged that the meshing step required to obtain a suitable geometry may take 90% of the time needed to obtain an engineering result. The objective of our work is to develop a technology to embed “dirty” geometries within a background mesh, which is then adapted to fit the requirements of the simulation. The technique employed results in a methodology, which is both robust and scalable on modern HPC hardware.

PRESENTATION PDF DOWNLOAD

Title: EXA2CT: Mining Chemical Space Annotation to tackle the Phenotypic Challenge of Pharma Industry

Speaker

  • Hugo Ceulemans, Janssen

Abstract
The trajectory from a biological concept to a drug available to patients is expensive and typically spans over a decade. Drug discovery starts by mapping a disease mapped to a scalable experiment in a test tube. This enables the screening libraries of chemicals for hits or active compounds, from which chemical starting points or leads are selected. These leads are then taken through a cascade of follow-up assays and animal models to optimize their potency on the intended protein targets implicated in disease, while controlling their activity on undesired targets associated with side effects. Finally, the compound is transferred to drug development, where the candidate drugs are tested in human subjects in three subsequent clinical phases. Still, the vast majority of candidates that enter drug development do not make it through to approval. One current trend to mitigate the high attrition rate is to do the initial screening in more complex, so-called phenotypic assays, which are believed to emulate the disease much better than biochemical assays, and that do not rely on the limiting knowledge of which targets are critical for effect. The phenotypic approach, however, presents challenges of their own: their throughput is lower, implying a need for more compact libraries. Secondly, many of the existing compound optimization processes require knowledge of the target. Both of these challenges can be addressed by improving the industries capabilities to predict the activities of chemical on not just the intended protein target, but on as many proteins and processes as possible. To this end, we propose scaled-up machine learning approaches that can mine extensive but heterogeneous information on biological activities of chemicals that is accessible to the industry, to learn to predict it comprehensively. Moreover, we believe computational approaches enable us to extract much more relevant primary information for these exercises from industry standard screens; for instance, by more extensive image analysis, feature selection and machine learning microscopy based screens. Finally, progress is being made in not only formulating predictions, but also quantifying the reliability of predictions, not just for each model for a certain target, but even for individual prediction of a given chemical at a given concentration on a given target.

PRESENTATION PDF DOWNLOAD

EESI2 Final Conference

The second European Exascale Software Initiative, EESI2, organized its final international conference on May 28 noon – May 29 noon, 2015 in Dublin, back-to-back with PRACEdays15 at the AVIVA Stadium. More information about the program can be found at the EESI website.

EESI2 is an EC-funded FP7 project. The main goals of EESI2 are to elaborate an evolutive European vision and roadmap and to propose recommendations to address the challenges of Extreme Data and Extreme Computing on the new generation of Exascale computers expected in 2020.

More than 120 experts have been involved during the last two years on this project and documents including recommendations have already been published.

The EESI2 project presented all its works, vision and recommendations to a large public in a final two days conference, on 28 – 29 May 2015 in Dublin. This major event was organized by SURFsara and funded by the European Commission. The public includes worldwide specialists, scientists, engineers, policy makers from different European member states and representatives of the European Commission.

The first day of the event focussed on the technical challenges and recommendations in the areas of “Tools and Programming Models”, “Ultra Scalable Algorithms” and “Data Centric Approaches”, each of them being underlined with the needs of specific user applications. Discussion on the technical aspects of the recommendations was facilitated through a panel at the end of the day.

The second day focussed on the international HPC ecosystem, including co-design and education aspects, and the way forward beyond EESI2. The EC will present its strategy and plans in the HPC area as well. The day ended with a panel discussion on how organizing in future the eco-system, the collaborations at international level and between scientists and industrials. Important stakeholders and the EC contributed to the panel.