Abstract : Numerous numerical software libraries have high quality implementations of efficient algorithms and thus
facilitate transfer of new algorithms developed in academia into scientific applications, including those in labs
and industry. Recently, use of these libraries by labs and industry has significantly increased, and the value of
these libraries has been made clear. Several libraries have enabled applications to migrate their simulation codes
to the newest exascale systems as well as improve performance and capabilities. We will overview library
activities undertaken to ensure value to applications and give examples demonstrating use of new algorithms
as well as capability and speed improvements.
[01911] Supporting Applications with the Chombo Framework
Format : Talk at Waseda University
Author(s) :
Daniel Francis Martin (Lawrence Berkeley National Laboratory)
Abstract : Many scientific and industrial applications solve systems of partial differential equations, and can often benefit from algorithmic strategies like adaptive mesh refinement $(AMR)$, higher-order mapped grids, and linear and nonlinear solvers. These are often difficult to implement accurately and efficiently; software frameworks leverage this effort across many applications. We present case studies demonstrating how the modular design of the Chombo software framework supports performant applications, which then feed improved capabilities back into Chombo.
[01773] Factorization based sparse solvers and preconditioners for robust solutions
Format : Talk at Waseda University
Author(s) :
Xiaoye Sherry Li (Lawrence Berkeley National Laboratory)
Abstract : Many high fidelity simulation and data analysis involve
large-scale multiphysics and multiscale modeling problems that
generate highly ill-conditioned and indefinite algebraic equations.
The factorization based algorithms are indispensible building blocks
in the solver stack to solve these numerically challenging problems.
We will highlight how factorizations and low-rank
approximate factorizations can be effectively used as standalone
direct solvers or as preconditioners for iterative solvers.
The focus will be on recent advances in SuperLU and STRUMPACK
targeting at exascale machines and applications.
[01830] Overview and Application Experiences with SUNDIALS
Format : Talk at Waseda University
Author(s) :
Carol Woodward (Lawrence Livermore National Laboratory)
Cody Balos (Lawrence Livermore National Lab)
David Gardner (Lawrence Livermore National Laboratory)
Daniel Reynolds (Southern Methodist University)
Abstract : The SUNDIALS library of time integrators and nonlinear solvers has recently increased its support for large-scale GPU-based systems through new data structures and solver package interfaces. This talk will overview the SUNDIALS packages and recently added capabilities then show results on newly deployed systems. Several applications have taken advantage of these capabilities to improve their time integration performance, and results from a selection of these, including combustion, phase field modeling, and cosmology, will be shown.
Prepared by LLNL under Contract DE-AC52-07NA27344. LLNL-ABS-844590.
[01837] Paving the road for efficient volume coupling with preCICE
Format : Talk at Waseda University
Author(s) :
David Schneider (University of Stuttgart)
Benjamin Uekermann (University of Stuttgart)
Abstract : preCICE is an open-source coupling library enabling partitioned multi-physics simulations of separated software packages. It offers methods for data communication, equation coupling, and data mapping. This talk presents new data mapping concepts in preCICE in terms of efficiency and accuracy, which enable users to apply preCICE for large-scale coupling scenarios. The relevance of these new data mapping concepts is demonstrated by showing user-provided example cases.
Abstract : This talk will give an overview of the Ginkgo library and highlight its features through several integrations. Ginkgo is a modern C++ library composed of numerical linear algebra algorithms which are optimized for multicore processors and Nvidia, AMD, and Intel GPUs. The use of sustainable software development principles allows the rapid development of cutting edge algorithms with high-quality interfaces. Among others, these are used in plasma simulation, cardiac electrophysiology, or CFD.
[01829] The Need of Ecosystems of Numerical Libraries for Applications
Format : Talk at Waseda University
Author(s) :
Ulrike Meier Yang (Lawrence Livermore National Laboratory)
Abstract : The emergence of heterogeneous computers with increasingly complex architectures necessitates continuous adaptation of software to take advantage of increased performance potential. Thus, the use of multiple mathematical libraries designed by expert mathematicians and software developers is crucial for application codes. Often, there exist interoperabilities between these libraries. So, as each library is ported to a new computer architecture, it is also important that these libraries continue to work together. This requires a healthy well-designed ecosystem. This talk will discuss the importance of a well-adjusted ecosystem of math libraries and its impact on applications.
[01650] Exascale-Ready Adaptive Mesh Refinement Applications with AMReX
Format : Talk at Waseda University
Author(s) :
Andrew Myers (LBNL)
Abstract : AMReX is a block-structured adaptive mesh refinement library that supports a variety of advanced GPU and CPU architectures. I will describe AMReX and its associated ecosystem of application codes, spanning scientific domains such as astrophysics, plasma physics, wind farm modeling, epidemiology, and more. I will demonstrate how AMReX enables these codes to implement novel modeling capabilities involving a range of computational motifs and to run efficiently on some of the largest supercomputers in the world.
[01703] On the Design and Performance of Exascale Applications using the Trilinos Solver Framework
Format : Talk at Waseda University
Author(s) :
Roger Pawlowski (Sandia National Laboratories)
Abstract : The Trilinos project supplies algorithms and enabling technologies for the solution of large-scale, complex multi-physics engineering and scientific problems on new and emerging high-performance computing architectures. Capabilities include performance portability abstractions, linear solvers, preconditioners, nonlinear solvers, time integration and finite element discretization tools. This talk will focus on recent achievements in developing exascale applications using Trilinos. We will show results from plasma physics simulations for electromagnetic environments and aerodynamics simulations for hypersonic reentry.
[01626] Experience with Exascale Applications using PETSc/TAO
Format : Talk at Waseda University
Author(s) :
Todd Munson (Argonne National Laboratory)
Abstract : In this presentation, we will discuss the extensions of the PETSc/TAO library to support the GPU-based Frontier and Aurora exascale systems and provide some available performance results. We will then provide some perspectives on using PETSc/TAO from an application point of view and conclude with thoughts on future needs to support next generation architectures.
[01915] MFEM: Accelerating Efficient Solution of PDEs at Exascale
Format : Talk at Waseda University
Author(s) :
Tzanio Kolev (Lawrence Livermore National Laboratory)
Veselin Dobrev (Lawrence Livermore National Laboratory)
John Camier (Lawrence Livermore National Laboratory)
Vladimir Tomov (Lawrence Livermore National Laboratory)
Julian Andrej (Lawrence Livermore National Laboratory)
Will Pazner (Portland State University)
Abstract : Efficient exploitation of exascale architectures requires rethinking of the numerical algorithms used in PDE-based simulations to expose fine-grain parallelism and maximize arithmetic intensity. In this talk we present an overview of MFEM $\mathrm{(\href{https://mfem.org}{mfem.org})}$, a library for high-order finite element methods, which powers HPC applications in a wide variety of fields. We review recent advancements in MFEM’s discretization solver, and GPU-accelerated algorithms, and demonstrate their impact in several large-scale applications from the US Department of Energy.
[01456] Firedrake: Math to Supercomputer
Format : Talk at Waseda University
Author(s) :
Koki Sagiyama (Imperial College London)
David A. Ham (Imperial College London)
Abstract : Firedrake is an open-source Python package for solving PDEs using finite element methods.
Using the UFL language originally developed for the FEniCS project and other packages in the Firedrake's ecosystem,
Firedrake generates efficient finite element codes automatically from the math expressions provided by the users, allowing them to move from one idea to another quickly.
It also provides a transparent access to the linear/nonlinear solvers in the PETSc library.
Here, we will show some recent developments in Firedrake.
[01866] Scalability Study for Planewave DFT Solvers
Format : Talk at Waseda University
Author(s) :
Doru Thom Popovici (LBNL)
Mauro del Ben (LBNL)
Andrew Canning (LBNL)
Osni Marques (LBNL)
Abstract : Modern supercomputers vary in compute power and network capabilities. For example, Summit and Frontier make use of GPUs for accelerating computation, while relying on either a fat-tree or dragonfly topology for transferring data between the nodes. On the other hand, Fugaku has thousands of CPUs and uses a six-dimensional torus for communication. In this work, we want to study these differences in the context of scaling the eigenvalue solvers used in planewave DFT calculations. More specifically, we will focus on four algorithms meant to solve a nonlinear eigenvalue problem, namely Conjugate Gradient, RMM-DIIS, Jacobi Davidson and Unconstrained. We will show that for each algorithm different considerations must be taken when parallelizing the computation. We will provide proxy applications for each algorithm, and we will provide a thorough analysis of each code on some of the state-of-the-art supercomputers. We will emphasize that systematic approaches can be derived to guide the parallelization such that the computation can effectively use the compute and network resources.
