An Excalibur Design and Development Working Group
An Excalibur Design and Development Working Group
Getting scientific applications ready for exascale is a huge undertaking. Very few of today’s codes can take full advantage of current Tier-1 systems, such as the 750,000 core ARCHER-2, let alone the millions of cores on heterogeneous systems that are anticipated with exascale. This workshop will explore the challenges facing the scientific computing community, and cover exascale readiness techniques such as performance portability, algorithmic fault tolerance, asynchronous programming, and improving scalability. The workshop will include contributions from experts from international exascale initiatives such as ECP in the USA and EuroHPC. It will also include input from leading vendors such as Cray/HPE, NVIDIA, AMD, Intel and Codeplay. Key questions will be: (i) how different simulation tools can be software-engineered for interoperability at scale; (ii) how future software engineering and new languages abstractions most effectively support the separation of concerns in development; and, (iii) how multi-physics simulations can optimally exploit heterogeneous hardware platforms.
Programming supercomputers remains challenging. Multiple levels of parallelism must be exploited in the core, on the compute node, and between nodes. Heterogeneous hardware with accelerators further complicate development and achieving performance portability. The waLBerla framework addresses these challenges by providing building blocks for simulations on block-structured grids. Adaptive mesh refinement and load balancing are shown to scale. Meta-programming techniques are used to generate highly efficient code for CPUs and GPUs. The development workflow can thus start from the symbolic formulation of the problem. The steps of deriving the specific simulation method, simplifying the computation by symbolic manipulation, and eventually generating efficient code for various architectures is performed automatically. We will demonstrate the features of waLBerla with some multiphysics applications.
The story of writing codes in my lab for one algorithm, over and over, illustrates the sometimes tortuous progress of computational research. This tale speaks of the inherent tension between the students’ goals (one-time use, performance brag, thesis results) and the goals of reusable, sustainable software. ExaFMM is a new re-write of a high-performance implementation of the fast multipole method in C++. My students and a post-doc have written at least seven separate code bases for this algorithm, over more than 10 years, and we hope this one is the definitive one.
We take a holistic view of efficiently using human and compute resources via libraries and community building. Performance portability is a key ingredient, albeit with different meaning at different levels of the software stack. We discuss some traps and fallacies, as well as techniques and methods of evaluation that help to align incentives and improve communication.
I will explain Intel’s oneAPI initiative, including direct programming support for accelerators using SYCL 2020 and OpenMP 5. Special consideration will be give to programmers who do not want to port their code to ever platform using a different programming model or who need to maintain simulation codes for one or more decades.
We have been carrying out the FLAGSHIP 2020 to develop the Japanese next-generation flagship supercomputer, Post-K, named as “Fugaku” recently. In the project, we have designed a new Arm-SVE enabled processor, called A64FX, as well as the system including interconnect with the industry partner, Fujitsu. The processor is designed for energy-efficiency and sustained application performance. The “Fugaku” is being installed and scheduled to be put into operation for public service around 2021. In this talk, several features and software, programming tools of the “Fugaku” system will be presented as well as some preliminary performance results.
The range of computer architectures used in supercomputers today is growing in diversity. The Exascale machines due to be online by next year (2021) and the UK’s Tier-2 systems showcase the variety of architectures we need to think about: high-performance CPUs and GPUs from different hardware vendors. As HPC specialists, we need to write and maintain our codes to make sure they work efficiently on this wide and changing range of hardware. We want to write performance portable codes that run well everywhere. To do this we need robust ways to measure and evaluate performance portability, and we need to use write codes using programming models which are simultaneously performant and portable. This talk highlights some of the ongoing research into performance portability from the High Performance Computing Research Group at the University of Bristol.
Exascale systems will present application authors with a diverse set of programming models targeting a diverse set of hardware architectures. While many of these hardware architectures will be similar, the heterogeneity of each requires us to carefully consider our application designs and the technologies that will allow those applications to be portable between systems. In this talk, we’ll discuss the ecosystem of programming models that will be available in the coming years, including OpenMP, SYCL, Kokkos, RAJA, HIP, and CUDA, and the compiler technology required for each. In the space of programming models, there is a trade off between user control and the potential for compiler optimizations, and work exploring compiler optimizations for parallel programs will be highlighted. Finally, we’ll discuss auto-tuning and the potential advantages of using just-in-time compilation to maximize performance.
Open source, community-developed reusable scientific software represents a large and growing body of capabilities. Linux distributions, vendor software stacks and individual disciplined software product teams provide the scientific computing community with usable holistic software environments containing core open source software components. At the same time, new software capabilities make it into these distributions in a largely ad hoc fashion. The Extreme-scale Scientific Software Stack (E4S),first announced in November 2018, along with its community-organized scientific software development kits (SDKs), is a new community effort to create lightweight cross-team coordination of scientific software development, delivery and deployment and a set of support tools an processes targeted at improving scientific software quality via improved practices, policy, testing and coordination. E4S, which announced the release of Version 1.0 in November 2019, is an open architecture effort, welcoming teams that are developing technically compatible and high-quality products to participate in the community. E4S and the SDKs are sponsored by the US Department of Energy Exascale Computing Project (ECP), driven by our need to effectively develop, test, deliver and deploy our open source software products on next generation platform to the scientific community. In this presentation, we introduce E4S, discuss its design and implementation goals and show examples of success and challenges so far. We will also discuss our connection with other key community efforts we rely upon for our success and describe how collaboration around E4S can be realized.
Over recent years heterogeneous systems have become more popular across HPC systems, with over 100 supercomputers in the TOP500 incorporating GPUs or other accelerators. Such platforms have different performance characteristics and optimization requirements. In order to make the most of various accelerators, a programming model is required to support portability without losing performance. The SYCL programming model allows users to write heterogeneous programs using completely standard C++, and so developers have access to the power of C++ templates meta programming to develop a highly parameterized kernels. SYCL offers performance, programmability, and portability with extensive hardware support.
Developers can currently use SYCL to target a wide range of processors including those from Intel, Arm, NVidia and Renesas. Alongside this support, however, is a growing ecosystem of projects developed using SYCL that provide developers with optimized libraries for common operations and algorithms used in AI.
During this presentation, we will examine some of these projects. Show how these libraries can be used to offer portability and performance, and how these libraries can be used to accelerate complex HPC applications. Codeplay Software will discuss some of the available papers detailing the implementation, performance, and programmability of these libraries and frameworks.