Exascale-ready distributed and accelerated dense Numerical Linear Algebra (ExaNLA)

→ Europe/Berlin
Rotunda, Building 16.4 (Juelich Supercomputing Center)

Rotunda, Building 16.4

Juelich Supercomputing Center

Wilhelm Johnen strasse 52425 Juelich Germany
Edoardo Di Napoli (Forschungszentrum Juelich -- JSC), Xinzhe Wu (Forschungszentrum Juelich -- JSC)
Description

ExaNLA is a hands-on workshop named after a project aiming at creating a numerical linear algebra library for selected kernels capable of executing at scale on JUPITER and similarly massively parallel accelerated architectures.

Website

ExaNLA workshop page

PROGRAM
The objective of such workshop is multifold:

  1. Engage with the Community of Expert Users: Conduct an in-depth consultation process with domain experts to identify the essential core functionalities required by application codes, ensuring that the ExaNLA library meets the most pressing needs of its primary users.
  2. Initiate Brainstorming Sessions: Begin a series of collaborative discussions to address the needs and challenges in developing the new ExaNLA library.
    • Programming Languages: Evaluate the suitability of different programming languages for developing a high-performance and scalable ExaNLA library.
    • Parallel Programming Models: Explore various parallel programming models that can optimize the library’s performance on modern architectures.
    • Interfaces: Consider user-friendly and versatile interfaces that facilitate easy integration with existing applications and workflows.
    • Templating: Discuss the use of templating to enhance code reuse, flexibility, and maintainability, allowing for easier adaptation to different computational environments and various range of accuracy. 
    • Fault Tolerance: Design and develop an efficient fault tolerance middleware for the ExaNLA library, and explore options for integrating a checkpointing mechanism into it.
  3. Designing for Performance and Adaptability: Explore the feasibility of incorporating automatic mechanisms within the library to enhance performance and adaptability.
    • Automatic Symmetry Discovery and Algorithm Selection: Implement features that can automatically detect symmetries in matrices and select the most appropriate algorithms, optimizing computational efficiency. 
    • AI-Driven Optimization: Utilize AI techniques to optimize and select platform-specific kernels dynamically, ensuring that the library can leverage the best available hardware resources.
  4. Establish a Robust RSE Framework: Set up a research software engineering (RSE) framework based on Agile methodologies to ensure a streamlined and efficient development process. This framework incorporates: 
    • Continuous Integration (CI): Regularly integrate code changes to maintain software stability and identify issues early.
    • Continuous Development (CD): Enable ongoing software enhancements and feature additions.
    • Continuous Benchmarking (CB): Implement a process for regular performance testing to ensure the library meets its speed and scalability targets. 
    • Test Driven Development (TDD): Adopt a TDD approach to ensure code reliability and robustness, with tests guiding development from the outset.

  5. Application-driven benchmarking: Perpetual execution of application-driven benchmarks, including comparison against evolving existing libraries.

PARTICIPATION
Attendance to the event is by invitation only. The registration is moderated and it will be confirmed shortly after submission. It is expected to have participants from several disciplines spanning from application fields (e.g., physics, engineering, chemistry), to computer scientists, parallel programming experts, applied mathematicians and software engineers.

Participants
  • Alfredo Buttari
  • Alfredo Correa
  • Andrea Ferretti
  • Andreas Herten
  • Andreas Kleefeld
  • Ben Hourahine
  • Carolin Penke
  • Daniele Varsano
  • Davor Davidovic
  • Emmanuel Agullo
  • Fabio Affinito
  • Fabrizio Ferrari Ruffino
  • Gustavo Ramirez-Hidalgo
  • Hartwig Anzt
  • Hemanth Kolla
  • Ivo Kabadshow
  • Jayesh Badwaik
  • Jose E. Roman
  • Luigi Genovese
  • Maria Barrios Sazo
  • Mark Gates
  • Matteo Rizzi
  • Matthias Bollhöfer
  • Nur Aiman Fadel
  • Paolo Bientinesi
  • Petr Karpov
  • Pietro Delugas
  • Piotr Luszczek
  • Robert Speck
  • Stepan Nassyr
  • Thomas Kühne
  • Tommaso Benacchio
  • Toshiyuki Imamura
  • Volker Blum
  • +5
    • 8:30 AM → 9:00 AM
      Registration and welcome
      • 8:30 AM
        Welcome to JSC 15m
        Speaker: Robert Speck
      • 8:45 AM
        Goals and expectations 15m
        Speaker: Dr Edoardo Di Napoli (Forschungszentrum Juelich -- JSC)
    • 9:00 AM → 10:30 AM
      Application-driven functionalities: Survey of functionalities

      In depth brainstorming and consultation process to determine the most prominent functionalities needed by the communities of users and developers of application codes

      • 9:00 AM
        Chair opening 20m
        Speaker: Thomas Kühne
      • 9:20 AM
        Large scale semi-empirical electronic structure 5m
        Speaker: Ben Hourahine
      • 9:25 AM
        Linear algebra needs for Green's function methods as implemented in the Yambo code 10m
        Speaker: Andrea Ferretti
      • 9:35 AM
        Non-linear eigenvalue problems arising in acoustic scattering 10m
        Speaker: Andreas Kleefeld
      • 9:45 AM
        Recurrent linear algebra operations in tensor network algorithms and associated challenges 10m
        Speaker: Matteo Rizzi
      • 9:55 AM
        Optimized preconditioners for linear solvers in numerical weather prediction models 5m
        Speaker: Tommaso Benacchio
      • 10:00 AM
        Experience from application-driven problems in Numerical Linear Algebra 5m
        Speaker: Matthias Bollhöfer
      • 10:05 AM
        Eigensolver for electronic structure theory: ELPA experience 10m
        Speaker: Peter Karpov
    • 10:30 AM → 11:00 AM
      Coffee/Tea break 30m
    • 11:00 AM → 12:30 PM
      Application-driven functionalities: Selection of kernels

      In depth brainstorming and consultation process to determine the most prominent functionalities needed by the communities of users and developers of application codes

      • 11:00 AM
        Chair opening 20m
        Speaker: Luigi Genovese
      • 11:20 AM
        Core operations for electronic structure and Maxwell equations 5m
        Speaker: Ben Hourahine
      • 11:25 AM
        Dense needs arising from sparse applications 5m
        Speaker: Jose E. Roman
      • 11:30 AM
        Dense iterative eigensolvers need specialized kernels 10m
        Speaker: Dr Edoardo Di Napoli (Forschungszentrum Juelich -- JSC)
      • 11:40 AM
        Some questions 5m
        Speaker: Paolo Bientinesi
    • 12:30 PM → 1:30 PM
      Lunch break 1h
    • 1:30 PM → 3:00 PM
      Technical brainstorming sessions: Programming language
      • 1:30 PM
        Chair opening 20m
        Speaker: Toshiyuki Imamura
      • 1:50 PM
        PROGRAMMING LANGUAGES: some pros and cons 5m
        Speaker: Fabio Affinito
      • 1:55 PM
        Programming language selection, some thoughts 3m
        Speaker: Raffaele Solca`
      • 1:58 PM
        Meta-Programming in C++: Boosting Modularity or Breeding Bloat? 10m
        Speaker: Xinzhe Wu (Jülich Supercomputing Centre)
      • 2:08 PM
        Interoperability with Fortran 5m
        Speaker: Pietro Delugas
      • 2:13 PM
        Lessons learnt from porting a Fortran code to modern C++ 5m
        Speaker: Ivo Kabadshow
    • 3:00 PM → 3:30 PM
      Coffe/tea break 30m
    • 3:30 PM → 5:06 PM
      Technical brainstorming sessions: Parallel programming models
      • 3:30 PM
        Chair opening 20m
        Speaker: Hartwig Anzt
      • 3:50 PM
        Current State of GPU Programming Models 10m
        Speaker: Andreas Herten (Forschungszentrum Jülich)
      • 4:00 PM
        Choosing PARALLEL PARADIGMS: a hard task(ing) 5m
        Speaker: Fabio Affinito
      • 4:05 PM
        Avoiding vendor lock while maintaining performance: One code multiple targets, is it possible to maintain performance? 3m
        Speaker: Davor Davidovic
      • 4:10 PM
        On the scalability of task-based programming models 8m
        Speaker: Alfredo Buttari
      • 4:18 PM
        How to deal with hierarchical parallelism, HW heterogeneity and maintainability 5m
        Speaker: Ivo Kabadshow
      • 4:23 PM
        Experiences with C++ std::exec 8m
        Speaker: Raffaele Solca`
      • 4:31 PM
        Performance and Portability for C++ on Heterogeneous HPC Systems 5m
        Speaker: Jayesh Badwaik
      • 4:36 PM
        Communication Layers for Multi-GPU Systems: CUDA-Aware MPI, UCX, or NCCL/RCCL? 5m
        Speaker: Xinzhe Wu (Jülich Supercomputing Centre)
      • 4:41 PM
        Interoperability with the hosting parallelization model, supporting data redistributions and transpositions 5m
        Speaker: Pietro Delugas
    • 9:00 AM → 10:30 AM
      Technical brainstorming sessions: Fault tolerance
      • 9:00 AM
        Chair opening: Fault-tolerance for distributed scientific HPC: Needs and Challenges 20m
        Speaker: Hemanth Kolla
      • 9:20 AM
        Challenges between novel hardware and legacy software 10m
        Speaker: Nur Fadel
      • 9:30 AM
        Designing fault-tolerance solvers for the Exascale: the example of FT-GCR 10m
        Speaker: Tommaso Benacchio
      • 9:40 AM
        Hard lessons of flipping networks on JUWELS-Booster: a wake-up call for fault tolerance 3m
        Speaker: Xinzhe Wu (Jülich Supercomputing Centre)
      • 9:43 AM
        The scale of the running time and number resources used versus the cost of checkpointing/recovery: risk tolerance or opportunity cost? 10m
        Speaker: Piotr Luszczek
    • 10:30 AM → 11:00 AM
      Coffee/Tea break 30m
    • 11:00 AM → 12:30 PM
      Technical brainstorming sessions: Interfaces and templating
      • 11:00 AM
        Chair opening 20m
        Speaker: Mark Gates
      • 11:20 AM
        Linear algebra interfaces in Quantum Expresso 5m
        Speaker: Fabrizio Ferrari Ruffino
      • 11:25 AM
        Sparse BLAS standard interface 10m
        Speaker: Hartwig Anzt
      • 11:35 AM
        The ELSI experience 5m
        Speaker: Volker Blum
      • 11:40 AM
        Value and referential semantics for multidimensional arrays 10m
        Speaker: Alfredo Correa
      • 11:50 AM
        Interfaces, data structures and modeling environments in NLA 5m
        Speaker: Matthias Bollhöfer
      • 11:55 AM
        Interfacing application software and NLA libraries: data structures and parallel execution 5m
        Speaker: Dr Edoardo Di Napoli (Forschungszentrum Juelich -- JSC)
    • 12:30 PM → 1:30 PM
      Lunch break 1h
    • 1:30 PM → 3:00 PM
      Designing for performance and adaptability: Automatic symmetry discovery and algorithm selection
      • 1:30 PM
        Chair opening 20m
        Speaker: Paolo Bientinesi
      • 1:50 PM
        Algorithm selection driven by shape and condition: example of ill-conditioned TS QR factorization 10m
        Speaker: Davor Davidovic
      • 2:00 PM
        Going beyond the black-box paradigm: an example of exploitable knowledge 5m
        Speaker: Dr Edoardo Di Napoli (Forschungszentrum Juelich -- JSC)
    • 3:00 PM → 3:30 PM
      Coffe/tea break 30m
    • 3:30 PM → 5:00 PM
      Designing for performance and adaptability: Architecture-driven optimizations
      • 3:30 PM
        Chair opening 20m
        Speaker: Piotr Luszczek
      • 3:50 PM
        An Overview of Current and Imminent HPC Hardware 10m
        Speaker: Andreas Herten (Forschungszentrum Jülich)
      • 4:00 PM
        Exploiting Low Precision Tensor Cores without Losing Precision 5m
        Speaker: William Dawson
      • 4:05 PM
        Decoupling memory precision from arithmetic (in-core) precision 5m
        Speaker: Hartwig Anzt
      • 4:10 PM
        How low can we go? Floating-point challenges in numerical methods 5m
        Speaker: Ivo Kabadshow
      • 4:15 PM
        Exploiting hardware capabilities with kernel generation 10m
        Speaker: Stepan Nassyr
      • 4:25 PM
        MLIR: a possible general way for domain scientists? 10m
        Speaker: Nur Fadel
      • 4:35 PM
        Devil is in the ...software stack: the experience of the Poisson Solver of BigDFT code in porting on different arctitectures 10m
        Speaker: Luigi Genovese
    • 7:00 PM → 9:00 PM
      Social event: Social Dinner

      Dinner at a nearby Restaurant

      • 7:00 PM
        El Toro Restaurant 2h

        Große Rurstrasse 34
        52428 Jülich
        https://maps.app.goo.gl/ZRPbL7VyEew1pxFP9

    • 9:00 AM → 11:00 AM
      Software development and benchmarking: CI/CD, Test-driven Development and continuous benchmarking
      • 9:00 AM
        Chair opening 20m
        Speaker: Robert Speck
      • 9:20 AM
        CI/CD: a perspective from the HPC point of view 5m
        Speaker: Fabio Affinito
      • 9:25 AM
        CI experiences with a semi-empirical electronic structure code (DFTB+) 10m
        Speaker: Ben Hourahine
      • 9:35 AM
        CI, the good and the bad 5m
        Speaker: Raffaele Solca`
      • 9:40 AM
        Sematic test coverage with code mutation 10m
        Speaker: Alfredo Correa
      • 9:50 AM
        Continuous Benchmarking: an example setup at JSC to track a code's performance over time via gitlab CI 10m
        Speaker: Dirk Brömmel
      • 10:00 AM
        How to span the benchmarking space 5m
        Speaker: Fabio Affinito
      • 10:05 AM
        On reporting communication or network overheads 10m
        Speaker: Aravind Sankaran
      • 10:15 AM
        exaCB: A Scalable Framework for Continuous Benchmarking on Supercomputers 10m
        Speaker: Jayesh Badwaik
      • 10:25 AM
        Best Practices and Their Requirements 5m
        Speaker: Christoph Conrads
    • 11:00 AM → 11:30 AM
      Coffe/tea break 30m
    • 11:30 AM → 12:30 PM
      Conclusions: Recap of sessions and next steps
    • 12:30 PM → 1:30 PM
      Lunch break 1h
    • 1:30 PM → 2:00 PM
      Registration and welcome: Farewell