Page d'accueil // Recherche // Highlights // Ten University projects granted early access to MeluXina

Ten University projects granted early access to MeluXina

twitter linkedin facebook email this page
Publié le lundi 07 juin 2021

Luxembourg’s first supercomputer MeluXina was officially launched on 7 June. MeluXina is a key element of Luxembourg’s data-driven innovation strategy, catering for the needs of companies, start-ups, public and research institutions.

Ten research projects at the University of Luxembourg have been awarded early access to MeluXina, Luxembourg’s new supercomputer. For one month starting mid-June, the 10 selected projects will be able to run large-scale experiments and test their software on the system before the actual start of operations of MeluXina. The awarded projects were evaluated based on the potential impact on society, economy and science, as well as their capacity to exploit the supercomputer resources.

The selection was announced by LuxProvide, the Luxembourg High-Performance Computing & High-Performance Data Analytics centre (HPC/HPDA). Prof. Pascal Bouvry of the University of Luxembourg became CEO of LuxProvide in 2020, a position he shares with Roger Lampach.

HPC/HPDA are important pillars in the University’s digital strategy. The awarded research projects rely on HPC/HPDA, they require important computational power, generate and consume large volume of data to support their development and testing. The early access to MeluXina offers the researchers an excellent opportunity to assess the scalability of the models investigated in research projects and run their prototypes in a production environment. In addition, memory and communication bandwidth are usually a bottleneck for running large-scale experiments. MeluXina, which is part of the EURO-HPC supercomputers network, contains state-of-the-art processors and a 200Ggbps Infiniband Interconnect that address these challenges.

The preliminary results will provide incentives for other researchers to run their projects in MeluXina, to establish public-private partnerships for academic and industrial investigations on advanced research. One of the short to mid-term needs to be fulfilled at national level is the required skills by the industry for exploiting the capacity and benefits offered by the new supercomputer. The University of Luxembourg possesses the required capacity and plays an important role in educating the next generation of computational scientists.

Research led at the University supports the national research and innovation strategy, and it contributes to positioning Luxembourg as a knowledge-based economy. In that context, the 10 chosen projects support existing pillars of excellence in research in priority areas of materials science, physics and biology, materials and simulation, computer science and ICT. The application domains in health and systems biomedicine, artificial intelligence and machine learning, security, digital twin, quantum computing, finance, space, data modeling and simulation contribute to the scientific and economic competitiveness of the Grand Duchy of Luxembourg.

In addition, several research projects are the outcome of collaborations within research groups and between research and industry. Collaborations include projects with the Luxembourg Institute of Health (LIH), the University of Cambridge, Rafinex sarl, the European Investment Bank (EIB), ELIXIR infrastructure, and the FEniCS project.

Selection of research projects

GigaSom

Researchers Dr. Christophe Trefois (LCSB/Uni.lu, Project Owner), Dr. Laurent Heirendt (LCSB/Uni.lu), Dr. Miroslav Kratochvil (LCSB/Uni.lu).

GigaSOM.jl is a Julia package for clustering data from flow and mass cytometry. It is developed because the existing methods could not easily contain the datasets that are produced now. The motivation for the whole project came from Luxembourg Institute of Health (LIH) who use the software regularly now. The research group collaborates with Dr. Markus Ollert from LIH.

Dr. Kratochvil, explains: “It takes high-dimensional measurements of huge amounts of single cell that are "colored" by many (around 20 to 60) specific fluorescent dyes or heavy-ion markers, usually millions of cells per sample, and simplifies the dataset to a description of significant groups that can be identified within the cells. That is much easier for the experts to view and work with. For example, if you put (properly stained) blood into the cytometer, this allows you to identify all red & white blood cell types and subtypes, and very precisely count how much of each was found in the sample. This is indispensable e.g., for diagnosis and monitoring of leukemia in patients.

Application Domain Unsupervised clustering, dimensionality reduction and visualisation of multidimensional point clouds using the self-organising maps. The implementation is designed to work on many computers at once, because the current datasets may very easily not fit any single computer.

Aim, benefits for the project Dr. Kratochvil explains: “It is expected that the benchmarking will either assure us that the scalability is sufficient for MeluXina-style environments or point out the precise part of the software that we need to tune. It is also aimed to test several related methods (mainly the GPU versions of the same algorithm) to see whether the throughput cannot be improved even more. In short, any benchmarks on new hardware make the package more future-proof, one way or another.” Dr. Kratochvil said that “It gives us a great opportunity to test the scalability of the approach. Mainly, if the data distribution primitives will scale to say 10x or 100x larger computer clusters than what we tried so far (or if there is a problem that will surface), and if the speedups and efficiency can be maintained to larger datasets. In short, we want to make sure that we're ready for future huge datasets.”

Impact Dr. Kratochvil mentioned “In case the users of GigaSOM would like to run it on the new supercomputer, they will have some assurance that the application is "compatible", and they won't need to solve various installation and portability problems”, “GigaSOM is routinely used for diagnosing and evaluating human samples, which (in turn) partially contributes to answering many complicated research questions in immunology and oncology”.

The participation of Dr. Kratochvil on the project is an outcome of a cooperation within the ELIXIR infrastructure.

Assessing pre-exascale performance of the FEniCS Project Finite Element Software

Researchers Dr. Jack Hale (Uni.lu, Project Owner), Prof. Stéphane Bordas (Uni.lu), Mr. Michal Habera (Uni.lu), Mr. Martin Rehor (Rafinex sarl), Dr. Chris Richardson (University of Cambridge), Prof. Garth Wells (University of Cambridge), Prof. Andreas Zilian (Uni.lu), Dr. Xavier Besseron (Uni.lu).

The FEniCS Project finite element software is a computing platform for quickly translating scientific models into efficient finite element simulations. It has been used to develop robust and scalable finite element solvers for challenging problems in diverse application areas including Physics, Mathematics, Engineering and Biology.

The project brings together experts at the University of Luxembourg, University of Cambridge and Rafinex Sarl. It proposes to assess the scalability of four state-of-the-art finite element solvers implemented in the FEniCS Project with strong relevance to real problems in Science and Engineering. Strong and weak scaling tests will be carried out on MeluXina using up to 25600 processes.

Application Domain The linear elasticity problem; describes the elastic deformation of a solid body under the action of external loads. This type of analysis is common in engineering design. The FEniCS Project is used in Rafinex Sarl’s unique stochastic topology optimisation software, where thousands of similar elasticity models must be run in parallel to solve each problem.

This model can be used to examine e.g. the formation of Rayleigh-Bérnard convective cells, where characteristic rotational flows form due to density changes caused by heating and cooling of a fluid. Models of this type have applications in areas with complex multi-physics coupling including geophysics and engineering heat transfer

Aim, benefits for the project Dr. Hale expresses that “We have already demonstrated good strong and weak scaling of the FEniCS Project to 1000s of processes but performance at large process counts (>10000s) is still relatively unexplored. Memory and communication bandwidth are usually the bottleneck for our solvers. MeluXina contains state-of-the-art AMD EPYC Rome processors and a 200Gbps Infiniband Interconnect that can help us overcome these challenges.

“These results can demonstrate the capabilities of the FEniCS Project with a view to supporting proposals on pre-exascale computing to upcoming PRACE, EU and joint FNR/UKRI (INTER) calls. The use of MeluXina in an early access phase will give us a unique opportunity to test the scalability of internal data structures in the FEniCS Project to large numbers of processes.”

Impact Dr. Hale explains: “Positive results would show to our local industrial partners such as Rafinex Sarl that MeluXina is a cost-effective platform for delivering their FEniCS Project-based simulations to clients.”.

Large Scale Performance Study of XDEM

Researchers Dr. Xavier Besseron (Uni.lu, Project Owner), Prof. Bernhard Peters (Uni.lu), Dr. Jack Hale (Uni.lu).

The Extended Discrete Element Method (XDEM) is a multi-physics numerical tool developed at the University of Luxembourg for the simulation of the dynamic and thermodynamic of granular matter or particles. With many technical applications widespread in the Luxembourg’s industry, realistic simulations usually require a large number of particles and a long simulation time, and an important computing time. For that reason, XDEM is designed for high performance execution and supports parallel execution with MPI and OpenMP.

This project proposes to assess the performance of the XDEM software targeting parallel executions on 10,000 cores. Multiple software stacks, load-balancing policies, and thread configurations will be investigated in order to achieve the best performance on the new cutting-edge supercomputer MeluXina, and demonstrate the capability of the platform to simulate complex industrial problems.

Figure 1: Discharge of particles from hopper using a rotating chute. Thanks to parallelisation, millions of particles can be simulated in order to study, among others, the segregation of the minerals with different sizes.

Application Domain Discrete Element Method simulation or more generally multi-physics simulations.

Aim, benefits for the project Dr. Besseron mentions: “It aims to assess performance of our numerical simulation tool (XDEM) at large scale (MPI+OpenMP), validate our load-balancing policies, get a firsthand on the new cutting edge MeluXina platform.” Dr. Besseron indicated that the benefits for the project are testing this new platform – results at large scale.

Impact Dr. Besseron: “On the scientific aspect, this study will demonstrate the large-scale performance of the XDEM software, with a special focus on the original load-balancing policies and dynamic load-balancing specially designed for particles. XDEM has multi-physics applications such as biomass furnace, blast furnace and additive manufacturing [1] that are ubiquitous in Luxembourg’s industry. Efficient parallel numerical methods and analysis tools provide more detailed results, permit faster technological development and innovation and constitute an economic advantage”.

Dr. Besseron also indicated that “the goal of this project is not to run a complete real-life problem, but instead study how it scales-up.

Digital Twin of a Biomass Furnace

Researchers Prof. Bernhard Peters (Uni.lu, Project Owner), Dr. Xavier Besseron (Uni.lu).

Biomass as a renewable energy source continues to grow in popularity to reduce fossil fuel consumption for environmental and economic benefits. Biomass as a solid fuel with considerable amount of moisture and non-homogeneous properties, is difficult to be used in many applications. Combustion and gasification of biomass is the most promising road to follow, however, the processes are complex because of the involved physical and chemical processes on various time and length scales including heating up, drying, pyrolysis, oxidation of gaseous pyrolysis products, char formation and gasification. Therefore, a Digital Twin allows a shift from current empirical-based practice to an advanced multiphysics simulation technology including multi-physical models on different length-scales to mirror accurately the state of furnace which is not feasible through experimental measurements due to the inaccessibility of the packed bed of biomass and the hostile environment. Thus, this process is regarded to be the first step in a chain for innovative and functional design and determines significantly the quality of the furnace and its behaviour.

Application Domain The requirements of today’s fiercely competitive markets require better products at shorter innovation cycles. Therefore, a truly new and disruptive smart design paradigm is urgently needed for which the visionary objective is to advance high-performance computing technology for smart prototyping of biomass furnace. The XDEM simulation environment allows creating a multi-physics digital twin of a biomass furnace representing the thermal conversion process of both the moving pellets bed and the freeboard. Hence, expensive and time-consuming physical prototyping is avoided and a large number of different parameters is easily investigated. In addition, the effect of isolated single parameters is now traceable while other parameters remain constant (ceteris paribus) which physical prototypes simply do not allow. A thorough analysis of predicted results uncovers causal and hidden relationships that is otherwise not possible in an experimental framework. This newly gained knowledge serves as a basis for developing cutting-edge technology, is inexpensive and shortens the innovation cycles, which altogether provides a strong advantage in a competitive market.

Impact An additional and much broader impact is is generated by strengthening the links between public and private sectors through private-public-partnerships is one of the cornerstones of Luxembourg’s path towards Smart Specialisation and as laid out by the OECD report. It aims at intensifying research, technological development and innovation (RDI) activates by concentrating on Key Enabling Technologies (KETs) such as sustainability. The proposed digital twin concept has a significant impact on processing of biomass and is considered as a crucial step along the processing chain to the desired high quality biomass furnace encompassing functionality and durability. Creating a digital twin helps to unveil the underlying physics of biomass conversion, and thus, gaining a deepened understanding. The latter enables engineers to design improved reactors and operate them at more favourable conditions with a higher output at reduced costs contributing to a resource efficient Europe. The Digital Twin can also predict responses of the biomass furnace to safety critical events and uncover previously unknown issues before they become critical and thus targets also the societal aspect of safe and reliable processes.

Scalable A.I. Recommitment System

Researchers Prof. Pascal Bouvry (Uni.lu-LuxProvide, Project Owner), Dr. Emmanuel Kieffer (Uni.lu), Dr. Frederic Pinel (Uni.lu).

The University of Luxembourg and the European Investment Bank (EIB) through the STAREBEI programme are working together to encourage private equity partners to invest in innovative and sustainable technologies. The funded research project “Sustainable and Trustworthy Artificial Intelligence Recommitment System (STAIRS) ” proposes an innovative approach to generate efficient recommitment strategies to guide institutional investors with the aid of AI-based algorithms.

The automatic generation of strategies relies on a tremendous number of simulated private equity portfolios to tackle different market conditions. The multi-objective nature of the recommitment needs (ESG, liquidity, etc.) increases the computational requirements to provide robust and trustworthy strategies. Intensive simulations can only be sustained using a scalable and distributed A.I. algorithm taking advantage of High-Performance Computing platforms. This project would demonstrate the potential of STAIRS at large scale while revealing HPC to the financial world.

Application Domain Private Equity, Artificial Intelligence, Fintech.

Aim, benefits for the project Dr. Kieffer mentions: “To support all the development and tests, this project will strongly rely on High Performance Computing (HPC) to cope with the computing power requested by such an AI-based system. The use of MeluXina will be decisive to push back the frontiers of achievable while reducing tremendously the time needed to provide satisfying solutions”. Dr. Kieffer also stated that “achieving and maintaining high allocation to private equity and keeping allocations at the targeted level through recommitment strategies is a complex task and needs to be balanced against the risk of becoming a defaulting investor. When looking at recommitments we are quickly faced with a combinatorial explosion of the solution space, rendering explicit enumeration impossible. Strategies need to be evaluated through numerous and time-consuming simulations. This cannot be achieved on a classical computing architecture.

Impact The rise of Environmental, Social, and Governance (ESG) factors has been one of the major changes for private equity partners. ESG considerations have redesigned the standards of due diligence and add new objectives on top of financial statements and growth plans. Building private equity portfolios remains a real challenge for limited partners investors with heavy consequences on ESG/Sustainable investments. This lack of guidance is certainly the main barrier to overcome in order to give confidence to investors and encourage investing in innovative and sustainable technologies. Policy makers have tasked institutional investors such as the European Investment Bank (EIB) to invest in a sustainable future for all. Nevertheless, the different objectives, levels of risk aversion, ESG exposure and time-horizons are subject to complex constraints and trade-offs. Under such circumstances, there is a real need to design guidance mechanisms to leverage private equity responsible investments.”

Additional information can be found on the project's website

Pure GPU Constraint Solver

Researchers Prof. Pascal Bourvy (Uni.lu-LuxProvide, Project Owner), Dr. Pierre Talbot (SnT/Uni.lu), Dr. Frederic Pinel (Uni.lu)

The GPU Constraint Solver project aims to design a novel software architecture for solving constraint problems, a general method for many optimisation problems. Constraint solvers are very compute intensive, parallel machines offer a great opportunity to improve the solvers' performance. The novel architecture exploits mathematical properties to guarantee correct results on parallel machines.

Application Domain The overall domain is constraint solving, specifically the design and development of parallel constraint solvers. Our approach is based on recent theoretical results, from lattice theory (which provides provable properties for shared memory access), namely concurrent propagators on shared memory without synchronisation.

Aim, benefits for the project Dr. Pinel explains: “Access to MeluXina provides us with powerful parallel hardware, which can validate, and stretch, our novel design.” “To MeluXina, this offers a new model for parallel application design, and offers access to a new, open-source, constraint solver, developed from scratch in Luxembourg.

Impact The essential outcome is to test the validity of the chosen theoretical approach to parallel constraint solvers. This work is set in an overall research line, planned in a FNR CORE proposal, to be submitted in April 2021. The research line addresses the suitability of lattice theory to other parallel programming models (beyond constraint solvers), and even to define a new general purpose parallel model. Results from this scalability study will be very useful in the course of this CORE project.

Exploring the chemical space of drug-like molecules

Researchers Prof. Alexandre Tkatchenko (Uni.lu, Project Owner), Dr. Igor Poltavsky (Uni.lu), Artem Kokorin (Uni.lu), Dr. Leonardo Medrano Sandonas (Uni.lu)

Exploration of the chemical space of molecules with data-driven approaches has been inspiring countless academic and industrial initiatives for the design of novel drugs, antivirals, catalysts, battery materials, and in general chemicals with desired characteristics, a process which previously was largely driven by chemical intuition or serendipitous discoveries. Thus, this project aims to generate an extensive database containing reliable physicochemical property data of large drug-like molecules with relevance in biological and pharmaceutical applications. These data will be obtained by using highly-accurate quantum mechanical methods which require large number of computational infrastructures and hours of calculations. Machine learning (ML) models for the prediction of electronic and thermodynamic observables will also be trained employing the state-of-the-art ML frameworks developed in my research group. We expect that the outcomes of this project would boost the exploration and understanding of the complex chemical space of quantum mechanical properties for large molecules, contributing to the success of diverse chemistry decision-making processes.

Application Domain Dr. Medrano Sandonas said: “This project goes beyond pure scientific interests resulting in an enormous amount of pharmaceutical, biological, material science, and engineering applications.

Aim, and interest for the project Dr. Medrano Sandonas explains: “We are currently working on the development of novel machine learning (ML) methods for a faster and deeper exploration of the chemical space of quantum mechanical properties for large drug-like molecules. However, we are performing the benchmarks using the information of small molecules, which are very helpful for having a good understanding of our methods, but this limits the transferability of the outcomes to gain insights into the properties of, e.g., protein-ligand systems and molecules with pharmaceutical relevance. Thus, this early access to MeluXina supercomputer will give us the opportunity to complete the generation of a database of quantum mechanical properties for large drug-like molecules and, hence, to perform adequate benchmarks of our ML methods.

Benefits for the projectThis early access to MeluXina supercomputer will accelerate the generation of reliable quantum mechanical property data of large drug-like molecules with relevance in biological and pharmaceutical applications, which currently is inexistent. These data will help us to optimise our ML frameworks for the prediction of electronic and thermodynamic observables of large systems, providing a highly reliable basis for numerous applications for both scientific and industrial purposes. The outcomes of this project will also benefit the development of doctoral thesis and future research projects.

Additionally, as part of this project, we will carry out performance tests of the CPU, GPU, and hybrid CPU/GPU architectures in the MeluXina supercomputer. The quantum mechanical calculations will be performed only in CPUs, while our state-of-the-art ML frameworks operate in homogeneous and hybrid architectures. These tests will help other researchers for writing future proposals requesting computational resources in MeluXina.”

Impact Dr. Medrano Sandonas adds: “This project will contribute to the generation of one of the first databases of quantum mechanical properties for large drug-like molecules which could be used for the development of ML-assisted chemical space exploration tools for the discovery of chemicals with a desired combination of properties for a given application. Hence, these data will be the basis of future academic and industrial investigations in the direction of rational design of chemical compounds.

 

About the HPC at the University of Luxembourg

The University of Luxembourg operates since 2007 an excellent and large academic HPC facility that is considered as a one of the reference implementations at the Grand-Duchy of Luxembourg. The University’s HPC facility, managed by an expert team led by Professor Pascal Bouvry and Dr. Sébastien Varrette, offers a cutting-edge research infrastructure at national public research enhancing the digital competitiveness of the country, while serving as edge access to the Euro-HPC Luxembourg supercomputer.

Photo: © LuxProvide