PRACE Preparatory Access – 36th cut-off evaluation in March 2019

Find below the results of the 36th cut-off evaluation of 3 March 2019 for the PRACE Preparatory Access.

Projects from the following research areas:

 

Transport properties of graphene nanoribbon based devices

Project Name: Transport properties of graphene nanoribbon based devices
Project leader: Dr Carlo Antonio Pignedoli
Research field: Chemical Sciences and Materials
Resource awarded: 50000 core hours on SuperMUC-NG
Description

At Empa research is conducted in the field of carbon derived nanomaterials. In particular, after demonstration in 2010 of a bottom up approach for the fabrication of atomically precise graphene nanoribbons (GNRs) with different edge topologies, whose reliability and flexibility have been confirmed by its successful adoption in different laboratories around the world, the focus is now set on the fabrication of prototype field-effect transistors (FETs) based on GNRs. The research approach, based on a synergy between experiments and theory, has succeeded in the characterization of the electronic and optical properties of GNRs in gas phase and on metallic substrates. More recently, the possibility to fabricate a new family of GNRs was demonstrated. These structures are characterized by topologically protected states, whose electronic properties can be designed by varying the spacing between “extensions” added to the edges of a backbone. After the successful realization of the FET device, the next objective consists in characterizing, from a theoretical and an experimental point of view, the transport properties of different GNRs. In this project we join Empa’s capabilities in the fabrication of GNRs and in the theoretical investigation of their (electronic) structure with the expertise of Professor Mathieu Luisier from ETH Zurich in ab initio device simulation to study the transport properties of prototypical FETs based on an “all carbon” design. The outcome of our project will be a better understanding of the intrinsic factors that may limit the “current vs. voltage” characteristics of ultra-scaled GNR FETs, a study of their ultimate performance, and the design of transistors with optimized contact engineering. For prototypical devices we will compute, be means of Density Functional Theory (DFT), the equilibrium geometry of the ribbon/lead system, we will characterize electronically the ribbon/lead interface and extract the Schottky barrier and we will compute the phonons spectrum. In selected cases (i.e. for geometries that according to DFT calculations results more promising in terms of device performance) we will perform a higher level description of the ribbon/lead interface by means of GW calculations. Once stable device structures will be obtained, we will simulate their transport properties with a quantum transport solver, first in the ballistic regime or in the presence of structural/impurity defects, and, for the most promising components, with electron-phonon scattering. The key questions we want to answer within this projects are the following: What is the maximum current that GNR FETs with graphene leads can carry? How does the graphene-GNR contact interface affect the device performance? How does the Schottky barrier height (SBH) depend on the GNR type and on the leads material? Do the Schottky barrier height and injection efficiency calculated with quantum transport correlate? How do structural or impurity defects in leads or in the ribbon influence the transport properties? What is the effect of electron-phonon scattering on the transport properties of these low dimensional devices? Is self-heating important? How sensible are the transport properties to the ribbon/lead matching pattern?

top

Romulus-Zoom

Project Name: Romulus-Zoom
Project leader: Dr Benjamin Keller
Research field: Universe Sciences
Resource awarded: 100000 core hours on Marconi – KNL, 50000 cre hours on MareNostrum, 100000 core hours on Piz Daint
Description

We request preparatory access to perform scaling tests for a new generation of high resolution, zoom-in cosmological simulations. Our Romulus-Zoom project will ultimately make predictions for multimessenger observations of high-redshift supermassive black holes (SMBHs). It will transform our understanding of the relationship between the growth of SMBHs and their host galaxy. We wish to profile the performance of our codes on the following types of scenario: 1. Galaxies that undergo major mergers (1:4 mass ratio or greater) at z = 1−6 with final dynamical masses spanning the range Mvir = 10^10−10^12 Msun. These are likely hosts to SMBH mergers detectable by the future Laser Interferometer Space Antenna (LISA). Sampling the period of `cosmic noon’ where SMBH growth and star formation reach their peaks, these systems are ideal laboratories for multimessenger astrophysics. 2. Massive halos (Mvir ~10^12.5 Msun). With rich and varied merger histories and galaxies hosting SMBHs in the range probed by pulsar timing arrays (10^7-10^9 Msun), these are important laboratories for studying the connection between merger history and SMBH-galaxy co-evolution. Three such halos will be selected. 3. Milky Way (MW)-mass (Mvir ~10^11.7-10^12.2 Msun) halos. Combined with the above more massive halos, this suite will complete a sampling around the peak of the stellar mass-halo mass relation, where a transition is thought to occur from supernovae to SMBH feedback-regulated star formation. At a resolution orders of magnitude better than what was possible with Romulus25 (80pc force resolution, 16 pc hydro resolution, and 10^3 Msun baryon mass resolution), these simulations will implement the most realistic and predictive models for the ISM, supernovae feedback, and SMBH physics. A new implementation of SMBH feedback will utilize a model for thermal conduction and evaporation that, when combined with the detailed ISM model possible at such high resolution, will result in more realistic interactions between SMBHs and nearby multiphase gas. SMBH formation will depend on gas properties and the local radiation field from nearby star formation. SMBH dynamical evolution will be followed to ~150 pc separations, predicting where and when SMBH binaries form with unprecedented accuracy. All of these developments mean that accurate profiling is essential as a preparatory step to submitting a full PRACE proposal.

top

Excited-State Phenomena at Interfaces from Many-Body Perturbation Theory

Project Name: Excited-State Phenomena at Interfaces from Many-Body Perturbation Theory
Project leader: Dr. Sivan Refaely-Abramson
Research field: Chemical Sciences and Materials
Resource awarded: 50000 core hours on MareNostrum
Description

Designing new materials with optimized catalytic properties is of great importance for more sustainable energy production. Fundamental understanding of the processes taking place at the electrode-electrolyte interface is needed to set design rules and control pathways. Here we propose a combined experimental-computational project to gain a comprehensive picture of optimal catalytic surfaces for oxygen reduction by generating a mutual back-and-forth feedback between experimental results obtained with Scanning Tunneling Microscopy and X-Ray Photoelectron Spectroscopy studies of surfaces in different environments, and applying advanced many-body perturbation theory calculations to connect the results with dynamical bandstructure properties of the examined surfaces. The preparatory access will be used to examine the scalability of the main code used for these computations. Berkeley GW, on the MareNostrum machine.

top

UEABS_GMX_NAMD

Project Name: UEABS_GMX_NAMD
Project leader: Dr Dimitris Dellis
Research field: Chemical Sciences and Materials
Resource awarded: 100000 core hours on Marconi – KNL
Description

Gromacs and NAMD Molecular Dynamics packages will be benchmarked on the Tier-0 systems. Benchmark data sets are already published in prace web site. The results of performance, scaling as well as their hybrid parallelization behavior will be published in Prace 5IP deliverable : D7.5 Evaluation of Accelerated and Non-accelerated Benchmarks.

top

High resolution glacial climate conditions over the Alps: Regional Modeling

Project Name: High resolution glacial climate conditions over the Alps: Regional Modeling
Project leader: Dr. Jonathan Buzan
Research field: Earth System Sciences
Resource awarded: 100000 core hours on Marconi – KNL
Description

This production proposal summarizes the computational needs for the new project HicAp funded by the NAGRA (National Cooperative for the Disposal of Radioactive Waste, Switzerland). We will test both WRF and CESM configurations to optimize spin up and production simulation throughput. The overarching aim of the project is to assess a variety of past glacial states in order to predict the plausible range of possible climate conditions of a future glaciation of Switzerland. A deeper understanding about a future glaciation of Switzerland is needed to identify secure places for the disposal of radioactive waste in Switzerland as a potential thread is the over deepening of valleys during glacial times. Modeling and understanding climate conditions and changes on time scales up to 1 million years remains one of the grand challenges in climate sciences. For Switzerland, specific additional challenges have to be considered, e.g., it is an area with complex terrain. To gain precise knowledge of the driving mechanisms of the Alpine ice sheet this project will utilize a unique model chain of comprehensive global Earth system and regional climate models. With a set of sensitivity simulations, we plan to sample the parameter space of the main drivers of Atlantic European climate on these time scales, namely, orbital forcing and Northern Hemisphere ice sheets (mainly the Laurentide ice sheet). We will assess glacial climate states in the past on the global scale and in a very high horizontal resolution (2 km) over Switzerland enabling ice sheet and land surface modelers to use the meteorological data, in particular temperature and precipitation. Regional climate modeling at this resolution is new for paleoclimatology. Thus, several modeling development steps concerning, e.g., the surface boundary are needed. Finally, statistical techniques will be applied to the simulations in order to provide a tool applicable to future climate projections. The project is of high national importance and certainly at the forefront of paleoclimatology. Highly resolved climatic information suitable to simulate the ice sheet and glacier behavior over Switzerland during glacial times is the actionable product from this project. This pre-proposal project aims to maximize production capacity of two developed modeling frameworks: 1) Earth system models acting as large scale boundary conditions for 2) weather research and forecasting models, used to resolve atmospheric processes in high spatial resolution.

top

High resolution simulations of climate change and land use to project their impact in Peru and Kenya

Project Name: High resolution simulations of climate change and land use to project their impact in Peru and Kenya
Project leader: Dr Santos J. González-Rojí
Research field: Earth System Sciences
Resource awarded: 100000 core hours on Marconi – KNL
Description

This production proposal summarizes the computational needs for a new project developed at the Climate and Environmental Physics Division of the University of Bern (Switzerland). The objective of the project is to create a variety of highly resolved simulations to study the effect of future climate change and land use change over key vulnerable regions in Peru and Kenya. Modelling and understanding, and eventually credibly projecting, the impact of climate change on the freshwater supply is one of the main challenges in climate sciences. A particular focus are regions with complex topography where the population is steadily growing and where the land use changes according to their needs. To gain precise knowledge of the effect of these changes in the atmosphere, this project will utilize the regional downscaling model WRF (Weather Research and Forecasting). With a set of sensitivity simulations, we plan to investigate the impact of land use change, which help us identify the dominant drivers of changes in precipitation patterns and thus seasonal and annual water availability. The simulations will use large domains over Central Africa and Western South America in which a hierarchy of smaller grids at increasing resolution are embedded to reach convection-permitting scales. The innermost domains cover the regions around Mt. Kenya and Madre de Dios in Peru. We propose a suite of baseline simulations for this project: simulate the current climate using the CESM climate model as initial and boundary conditions, and the end of the 21st century using the mitigation scenario RCP2.6 and the “business-as-usual” scenario RCP8.5 for the global climate simulation that drives the WRF model. From these baseline simulations we can already assess the effect of climate change on temperature and precipitation around the study areas. The baseline simulations then are the starting points of sensitivity studies concerning land use changes. This will allow us to assess sensitivities to climate and land use changes and finally estimate future water availability in Peru and Kenya. As a first step, this pre-proposal project aims to maximize the production capacity of the WRF model by testing different configurations of it in both chosen regions.

top

Testing computational methods to a molecular understanding of flaviviruses proteins

Project Name: Testing computational methods to a molecular understanding of flaviviruses proteins
Project leader: Prof. CATHERINE ETCHEBEST
Research field: Biochemistry, Bioinformatics and Life sciences
Resource awarded: 50000 core hours on Joliot Curie – SKL
Description

This project aims at developing, applying and integrate computational approaches to better understand molecular mechanisms in general and also specifically for flaviviruses proteins. The recent Zika, Dengue and Yellow fever viruses epidemics have raised several health concerns which motivated a race to decipher their molecular basis than can provide insights in their control interventions. Diagnostic is still a difficult issue since the clinical symptoms are highly similar. The understanding of their common structural/dynamical and biomolecular interactions features and differences might suggest alternative strategies towards differential serological diagnostics and therapeutic intervention. Due to their immunogenicity, the primary focus of this study will be on the Zika and Dengue non-structural proteins 1 (NS1) protein. Some experimental evidences bring electrostatic interactions as a key trigger factor. This topic is highly challenging and extremely important in many biological processes. By means of constant-pH methods under development by us, we will map possible epitopes and investigate their interactions with antibodies at different physical chemical conditions. This requires large computational resources due to a) the electrostatic coupling between the ionizable protein groups; b) and the need to repeat a vast number of independent simulation runs at several different experimental conditions, protein systems and conformations. A diversity of protein conformation will be obtained by clustering structures from classical molecular dynamics trajectories. Results will be compared with available experimental data and predictions by other Bioinformatics tools that often do not take into account pH effects despite their critical influence in the systems. The present study also opens up perspectives to computationally design high specificity antibodies.

top

Convergence tests in Particle-in-cell simulations of laser-plasma interactions

Project Name: Convergence tests in Particle-in-cell simulations of laser-plasma interactions
Project leader: Dr Zsolt Lécz
Research field: Fundamental Physics
Resource awarded: 100000 core hours on Hazel Hen, 20000 core hours on JUWELS
Description

The aim of the project is to find the minimum required mesh resolution in a Particle-in-Cell simulation at which the results are physically reliable and converge. The results of laser-plasma interaction simulations can be altered significantly by changing the resolution by factor of 2.

top

The PIC view on electron acceleration by cosmic shocks

Project Name: The PIC view on electron acceleration by cosmic shocks
Project leader: Dr Franco vazza
Research field: Universe Sciences
Resource awarded: 100000 core hours on Marconi – KNL
Description

We will use new particle-in-cell (PIC) simulations to explore the physics of collisionless cosmic shocks as well as their efficiency in accelerating cosmic ray particles. Cosmological MHD/hydro simulations have been extensively used to predict the distribution of cosmic shocks around and within large-scale structures, following their matter build-up. However, the details of the shock structure are necessarily unresolved, since the shock transition occurs on microscopic plasma scales (typically, the proton Lamor radius). This affects our understanding of the astrophysical meaning of telescope observations from the gamma to the radio domain. Shocks in the intergalactic plasma are collisionless, i.e. they are governed by collective electromagnetic effects, rather than binary particle collisions (which are extremely rare, in the dilute gas of galaxy clusters). The most fundamental way of capturing the dynamics of collisionless plasmas is by means of PIC simulations. In this proposal, we plan to extract the relevant shock conditions from cosmological MHD/hydro simulations and study with PIC simulations (produced with the highly optimized Tristan-MP code) how such shocks can partition the flow energy between protons and electrons, and how they can accelerate a fraction of particles to high energies, resulting in a power-law distribution extending beyond the thermal Maxwellian. High-energy electrons accelerated at shocks will emit synchrotron photons, thus explaining the variety of structures seen in galaxy clusters in the radio band. Thanks to our PIC simulations, we will be able to predict the injection efficiency of electrons and protons at galaxy cluster shocks (i.e., the fraction of particles participating in the acceleration process) and the slope of the resulting power-law tail. Such quantities can only be reliably assessed by running the simulations to unprecedentedly long times with PRACE supercomputing resources.

top

Scalability testing of cosmological hydrodynamical simulations with Nyx in preparation for the DESI survey

Project Name: Scalability testing of cosmological hydrodynamical simulations with Nyx in preparation for the DESI survey
Project leader: Dr. Michael Walther
Research field: Universe Sciences
Resource awarded: 50000 core hours on Joliot Curie – KNL, 50000 core hours on Joliot Curie – SKL
Description

The Lyman-alpha forest has proven to be a major tool for cosmology. Obtaining cosmological constraints based on the data delivered from observations, however, relies on grids of high-resolution, large volume hydrodynamical simulations which were not feasible in the past. The typical simulation grids used in this field consist of simulations with 1024^3 particles/grid cells or less with individual larger simulations for convergence studies. The development of fast, highly scalable simulation codes in recent years as well as the rise in available computational resources brought grids of simulations consisting of at least 2048^3 if not 4096^3 particles/grid cells into reach. The goal of this proposal is to assess scalability of such simulations using the state of the art code Nyx in preparation for running a cosmological analysis simulation grid planned for PRACE Cycle 19. In detail we wish to obtain the accuracy of Lyman-alpha forest simulations with different resolutions as well as scalings of computational time and memory consumption with resolution and simulation volume requring 5 additional simulation runs. After testing the scalability of the code we can decide on the best compromise between the number of simulations run in our analysis grid and the size of each individual simulation. While the former allows for higher interpolation accuracy between simulations, the latter allows for higher accuracy in the individual runs. Finally, while we used the code before on Joliot Curie/SKL nodes and while it’s been optimized to work on KNL nodes at NERSC/Cori/KNL, so far it has not been run at Joliot Curie/KNL. We’d like to assess the relative performance of the code on this system to enable choosing the optimal architecture for the computation of our analysis grid. In total we require 50000 CPU-h (including overheads for testing) on Joliot Curie/SKL for our purposes. To decide between both architectures another 15000 CPU-h for KNL are required.

top

The early growth of supermassive black holes

Project Name: The early growth of supermassive black holes
Project leader: Dr Alessandro Lupi
Research field: Universe Sciences
Resource awarded: 50000 core hours on Joliot Curie – SKL
Description

Observations of high redshift quasars (z >~ 6) show that black holes (MBHs) as massive as 10^9 Msun were already in place when the Universe was less than 1 Gyr old. Recent theoretical and observational results suggest that these MBHs inhabit massive haloes (Mhalo >~ 10^12 Msun; Di Matteo et al. 2017), but they do not seem to live in very biased regions of the Universe (Uchiyama et al. 2018). The formation and early growth of these objects, and the interplay with their galaxy host is still an open question, and only recently numerical simulations have reached a resolution high enough to start to resolve the interstellar medium of these high-redshift massive galaxies in detail (Smidt et al. 2018, Lupi et al. 2019). Among the different mechanisms proposed to explain the large MBH masses observed at high redshift, there is the intermittent accretion above the Eddington limit (Madau et al. 2014, Lupi et al. 2016). The purpose of the proposed project is to investigate the early growth of MBHs hosted in z ~ 6 quasars and the interplay with the ISM via high resolution zoom-in cosmological simulations, to assess the co-evolution between the central MBH and its host, and the possible recurrence of super-Eddington accretion phases during the MBH early evolution. We will also assess what conditions are responsible for the triggering of these super-Eddington phases, and what is the impact of the resulting feedback on to the interstellar medium of the galaxy host. Our simulations accurately model the non-equilibrium chemistry of the gas (following both primordial species (H,He,H2; Lupi et al. 2018; Lupi et al. 2019) and the main metal species (C,O,Si, N, and Fe; Capelo et al. 2018; Lupi et al. in prep.), star formation and stellar feedback (via winds, radiation, and supernovae) and black hole accretion and feedback (via winds, jets and X-ray radiation; see, e.g. Regan et al. 2018). Thanks to current observations by ALMA, and future ones at rest-frame near-IR (after subtraction of the PSF) by the upcoming JWST mission, we will soon be able to study the hosts of high-z quasars with extreme detail our simulations will be crucial for the joint interpretation of these observations.

top

Scalability tests for a new load-balancing method for parallel combustion simulations using finite-rates chemistry

Project Name: Scalability tests for a new load-balancing method for parallel combustion simulations using finite-rates chemistry
Project leader: Dr Jordi Muela Castro
Research field: Engineering
Resource awarded: 50000 core hours on MareNostrum
Description

The development and assessment of an efficient parallelization method for the evaluation of the reaction rates in combustion simulations is investigated. Combustion simulations where the finite-rate chemistry model is employed have a heavy computational load. In these simulations, a transport equation for each one of the species present in the chemical reaction mechanism have to be solved, and the resulting system of equations is stiff. Stiff sets of equations are a kind of ordinary differential equations (ODEs) characterized for presenting numerical instabilities, unless a very small time-step is employed. This limited time-step is inherent for both explicit numerical schemes as well as classical implicit integration methods employing fixed point iteration. Therefore, advanced implicit methods must be applied to obtain accurate solutions using reasonable time-steps. Nonetheless, these well-suited implicit methods for stiff equations demand higher computational resources than explicit or classical implicit methods. Therefore, in the present work a new algorithm aimed to enhance the numerical performance of the time integration of stiff equation systems in parallel combustion simulations is investigated. The developed algorithm identifies the control volumes where the system of equations is stiff. Then, they are integrated using Gear’s method, while non-reactive control volumes, that do not present a stiff behaviour, are integrated explicitly. Since reactive zones are close to the front flame, in parallel simulations only a few processors where the flame lives present a large amount of cells with stiff equations, while the other processors far from the flame can be integrated fully explicitly, creating a heavy imbalance in the computations. In order to solve this issue, an efficient dynamic load-balancing method of the computational load has been developed, increasing noteworthy the computational performance of the simulations, and consequently, reducing significantly the computer time required to perform the numerical combustion studies.

top

Type B: Code development and optimization by the applicant (without PRACE support) (9)

High Order Discontinuous Galerkin Solver for DNS computation

Project Name: High Order Discontinuous Galerkin Solver for DNS computation
Project leader: Prof. Eusebio Valero
Research field: Engineering
Resource awarded: 100000 core hours on MareNostrum
Description

High order Discontinuous Galerkin methods are shown an increasing maturity in their application to fluid problems where accuracy is a must. In combination with multi-domain decompositions are able to tackle with any kind of complex geometry and their low dissipative and high accuracy (for the same mesh) properties make them the perfect candidate to the analysis of complex flow configurations where large unsteadiness and a variety of different scales are found. Additionally, h/p adaptivity can provide an extra advantage in the optimization of degrees of freedom, increasing the mesh or the polynomial order in the region of interest. However, the application of these methods to realistic problems requires an efficient parallelization of the solver and the appropriate use of the large databases generated in the simulations. In this work, the performance and scalability of the newly developed high order discontinuous galekin solver will be evaluated against two representative test cases: standard DNS of a channel compressible flow and the wavy wall channel flow. The tool solves the Tridimensional Laminar Navier Stokes equations using a multidomain approach. An external or “h” mesh defines the general feature of the geometry and an internal o “p” mesh defines the resolution or degree of the polynomial in each domain. The connection between different domain is made following a mortar methodology: continuous in fluxes and discontinuous in the states variables. Hanging nodes or different degrees of freedom in different connected domains are allowed but makes the parallelization more challenge. Boundary continuous are imposed weakly. The solver can manage, and it is recommended, a high order representation of the boundary surfaces. The solver has been validated in UPM cesvima cluster (http://www.cesvima.upm.es) in the Taylor Green Vortex test case, until a maximum of 256 cores and 10 million nodes. A skew-symmetric formulation of the convective fluxes makes the solver robust in the analysis of under-resolved turbulence problems.

top

Unveiling turbulence-radiation interactions on heterogeneous computer architectures

Project Name: Unveiling turbulence-radiation interactions on heterogeneous computer architectures
Project leader: Dr. Rene Pecnik
Research field: Engineering
Resource awarded: 100000 core hours on Piz Daint
Description

The goal of this project is to perform first of its kind numerical simulations of irradiated highly participating turbulent flows to unveil the truly multi-physics interactions of convective, conductive and radiative heat transfer for enabling future solar energy systems. We use an innovative approach that exploits current heterogeneous high performance computing facilities. In particular, we solve (1) the fluid phase with direct numerical simulations (DNS) of the Navier-Stokes equations and (2) the radiative transfer equation (RTE) with the highly accurate, but computationally expensive, photon Monte-Carlo (MC) method. A Monte Carlo scheme is able to solve the RTE virtually to an analytical accuracy. On the other hand, the computational cost associated to an accuracy which is compatible with the DNS description is prohibitive with standard technology. Therefore, the bottleneck of this coupling is the MC code. Fortunately, the MC method is embarrassingly parallel, enabling efficient use of graphical processing units (GPUs). The coupling can, therefore, be achieved by adopting a heterogeneous computing approach.

top

UEABS: Code_Saturne for PRACE 5iP WP7 – Extra access

Project Name: UEABS: Code_Saturne for PRACE 5iP WP7 – Extra access
Project leader: Dr Charles Moulinec
Research field:
Resource awarded: 200000 core hours on Marconi – KNL, 200000 core hours on Joliot Curie – KNL, 200000 core hours on Joliot Curie – SKL, 250000 core hours on Hazel Hen
Description

The purpose of this work is to test Code_Saturne at scale on various Tier0 systems with different architectures. This work is undertaken through PRACE for its UEABS.

top

Extending Alya’s physical models to GPU computing

Project Name: Extending Alya’s physical models to GPU computing
Project leader: Dr Guillaume Houzeaux
Research field: Engineering
Resource awarded: 100000 core hours on Piz Daint
Description

This project is developed in the context of the EXCELLERAT European Centre of Excellence for Engineering Applications (H2020, 2018-2021), and the Spanish funded project CHEST for the development of an HPC combustion and emissions model for the analysis of sustainable transport powertrains. (MICINN, 2018-2021). Combustion is one of the fields with high strategic importance and potential to fully exploit the future exascale systems. The simulation of real combustion systems demands for a framework capable of combining advanced turbulence, spray and combustion models. The simulations of this project will be performed with Alya; the high performance computational mechanics code developed in BSC. The physics solvable with the Alya system include incompressible/compressible flow, solid mechanics, chemistry, particle transport, heat transfer, turbulence modeling, electrical propagation, etc. In particular, this project is focused on the simulation of high-fidelity turbulent combustion of reacting flows that demand the utilization of complex physical models, better accuracy and larger temporal and spatial scales. Therefore, adapting the CFD codes to the new heterogeneous systems is fundamental to satisfy the constantly increasing computing demands. The current heterogeneous implementation of Alya consists of the two most common parts of the CFD codes: the matrix assembly and the iterative linear solvers. The objective of this project is to extend the heterogeneous implementation of Alya to the modules that are needed to perform combustion simulations. Two key points that we need to address for this goal are: a) Implementation of the Flamelet model to represent the turbulent flow interaction with the flames. The development of this model follows a similar strategy than the assembly of the matrix, and therefore it is planned to be solved by means of OpenACC. b) Implementation of the ELSA model for capturing the whole spray evolution, in particular including primary break-up and secondary atomization. Such a model involves a combination of a eulerian and lagrangian approach, and alternatives using CUDA will be explored.

top

KKR_spin_spiral

Project Name: KKR_spin_spiral
Project leader: Dr Eszter Simon
Research field: Fundamental Physics
Resource awarded: 50000 core hours on JUWELS
Description

In this project, the magnetic properties of thin films and multilayers are investigated using first principles calculations. These calculations give a proper tool for investigating and understanding the creation of different complex magnetic states in thin films. A large set of spin-configuration space of the system may be scanned by self-consistent spin-spiral calculations. Recently, we implemented the non-relativistic calculations of spin-spirals into the screened Korringha-Kohn Rostoker Green function code for layered structures. This new code in particular useful to study magnetic systems, where the role of the induced moments is relevant. Moreover, from the total energies obtained from self consistent spin-spiral calculations a spin-model of the corresponding system can be set up, going eventually beyond the usual two-spin interactions.

top

Fast and Fault Tolerant Linear Algebra

Project Name: Fast and Fault Tolerant Linear Algebra
Project leader: Dr. Oded Schwartz
Research field: Mathematics and Computer Sciences
Resource awarded: 100000 core hours on SuperMUC-NG
Description

The goals of the current project include high performance fault tolerant linear algebra algorithms. Goal I: Strassen’s matrix multiplication algorithm (1969) was followed by Winograd’s (1971) who improved it’s complexity by a constant factor. Many asymptotic improvements followed. Yet, they have done so at the cost of large, often gigantic, hidden constants. Consequently, for almost all feasible matrix dimensions Strassen-Winograd’s remains the fastest. The leading coefficient of Strassen-Winograd’s algorithm was believed to be optimal for matrix multiplication algorithms with 2 × 2 base case, due to a lower bound of Probert (1976). Surprisingly, we obtain a faster matrix multiplication algorithm, with the same base case size, the same asymptotic complexity, but with a smaller leading coefficient. To this end, we transform the input matrices to a different basis, which can be done fast. We discuss improvements in the communication costs, it’s effects on parallelization, and the extension of this method to other Strassen-like algorithms. We also prove a generalization of Probert’s lower bound that holds under change of basis. This shows that for matrix multiplication algorithms with a 2 × 2 base case, the leading coefficient of our algorithm cannot be further reduced, hence optimal. We intend to search, implement, test, and benchmark this and new related algorithms. Goal II: We propose high performance and general purpose fault resilience, based on composition of the computation-graphs of algorithms with error correcting codes. We intend to search, implement, test, and benchmark related algorithms.

top

EPEEC: European joint Effort toward a Highly Productive Programming Environment for Heterogeneous Exascale Computing

Project Name: EPEEC: European joint Effort toward a Highly Productive Programming Environment for Heterogeneous Exascale Computing
Project leader: Dr Antonio Peña
Research field: Mathematics and Computer Sciences
Resource awarded: 200000 core hours on Marconi – KNL, 100000 core hours on MareNostrum
Description

EPEEC’s main goal is to develop and deploy a production-ready parallel programming environment that turns upcoming overwhelmingly-heterogeneous exascale supercomputers into manageable platforms for domain application developers. The consortium will significantly advance and integrate existing state-of-the-art components based on European technology (programming models, runtime systems, and tools) with key features enabling 3 overarching objectives: high coding productivity, high performance, and energy awareness. An automatic generator of compiler directives will provide outstanding coding productivity from the very beginning of the application developing/porting process. Developers will be able to leverage either shared memory or distributed-shared memory programming flavours, and code in their preferred language: C, Fortran, or C++. EPEEC will ensure the composability and interoperability of its programming models and runtimes, which will incorporate specific features to handle data-intensive and extreme-data applications. Enhanced leading-edge performance tools will offer integral profiling, performance prediction, and visualisation of traces. Five applications representative of different relevant scientific domains will serve as part of a strong inter-disciplinary co-design approach and as technology demonstrators. EPEEC exploits results from past FET projects that led to the cutting-edge software components it builds upon, and pursues influencing the most relevant parallel programming standardisation bodies. The consortium is composed of European institutions and individuals with the highest expertise in their field, including not only leading research centres and universities but also SME/start-up companies, all of them recognised as high-tech innovators worldwide. EPEEC features the participation of young and high-potential researchers, and includes careful dissemination, exploitation, and public engagement plans. EPEEC aims at using PRACE-provided computing resources to optimise, scale and test our codes – we acknowledge that no production run will be made.

top

I/O optimasation of the Bifrost stellar atmosphere code

Project Name: I/O optimasation of the Bifrost stellar atmosphere code
Project leader: Dr Boris Gudiksen
Research field: Universe Sciences
Resource awarded: 200000 core hours on Marconi – KNL
Description

During this project, we will try to further develop the I/O from the already parallelised stellar atmosphere code Bifrost. Earlier awarded PRACE projects has revealed that the I/O from the massively parallel jobs on Marconi-KNL intermittently produced I/O errors. The source of these errors and ways to circumvent them will be explored.

top

Towards Exascale Bayesian Matrix Factorization

Project Name: Towards Exascale Bayesian Matrix Factorization
Project leader: Mr. Tom Vander Aa
Research field: Mathematics and Computer Sciences
Resource awarded: 250000 core hours on Hazel Hen
Description

Matrix factorization is a core machine learning technique for applications of collaborative filtering, such as recommender systems or drug discovery, where a data matrix Y is factorized into a product of two matrices, such that Y = X x W. The main task in such applications is to predict unobserved elements of a partially observed data matrix. Bayesian matrix factorization (BMF) formulates the matrix factorization task as a probabilistic model, with Bayesian inference conducted on the unknown matrices X and W. Advantages often associated with BMF include robustness to over-fitting and improved predictive accuracy, as well as flexible utilization of prior knowledge and side-data. Finally, for application domains such as drug discovery, the ability of the Bayesian approach to quantify uncertainty in predictions is of crucial importance.

top

Type C: Code development with support from experts from PRACE (1)

NEMO performance

Project Name: NEMO performance
Project leader: Mr Eric Maisonnave
Research field: Earth System Sciences
Resource awarded: 100000 core hours on MareNostrum
Description

NEMO for Nucleus for European Modelling of the Ocean is a state-of-the-art modelling framework for research activities and forecasting services in ocean and climate sciences, developed in a sustainable way by a European consortium. Recent performance studies [1] suggest that scalability is not the major well of future performance gain, neither horizontal resolution increase, whereas potentiality of extra developments accelerating cache access (horizontal domain tiling and single precision computations) is favourably evaluated. To better understand how cache access limits our performance, we propose, in collaboration with a supercomputing center that already investigated this topic (BSC), to further analyse our code behaviour and test new implementations [1] Maisonnave, E. and Masson, S., 2019: NEMO 4.0 performance: how to identify and reduce unnecessary communications , Technical Report, TR/CMGC/19/19, CECI, UMR CERFACS/CNRS No5318, France.

top

 

Type D: Optimisation work on a PRACE Tier-1 (0)