SEARCH · Search NASA
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Here, stencil codes are compute-intensive algorithms, in which data points arranged in a large grid are being recomputed repeatedly from the values of data points in a predefined neighborhood. This fixed neighborhood pattern is called a stencil. Stencils codes see wide-spread use in computing the discrete solutions of partial differential equations and systems composed of such equations.
This project investigated new techniques for compiling stencil and stencil-like computations. Stencils computations are commonly found in applications such as image processing, physical simulations, image processing, and machine learning. In recent years, many high-performance domain-specific languages (DSLs) have been proposed to optimize stencil computations. To leverage such DSLs, however, existing codes often need to be rewritten. Such rewriting is manual, labor intensive, and error-prone. To alleviate such issues, this project investigated the application of program synthesis and artificial learning techniques to enable stencil computations to automatically leverage new high-performance DSLs. Rather than constructing syntax driven rules, verified lifting uses program synthesis to search for a target code fragment to compile the given input code into. In addition, it also searches for a proof that validates how the found target code fragment preserves the semantics of the original input. Thus, the target code fragment is guaranteed to be semantically equivalent to the input.
In situ visualization and analysis is of increasing importance as the compute and I/O gap further widens with the advance to exascale capable computing. Yet, in situ methods impose resource constraints leading to the difficult task of balancing simulation code performance and the quality of analysis. Applications with tightly-coupled in situ visualization often achieve performance through spatial and temporal downsampling, a tradeoff which risks not capturing transient phenomena at sufficient fidelity. Determining a priori visualization parameters such as sampling rate is difficult without time and resource intensive experimentation. We present a method for reducing resource contention between in situ visualization and stencil codes on heterogeneous systems. This method permits full resolution replay through recording halos and the communication-free reconstruction of interior values uncoupled from the main simulation. We apply this method in the computational fluid dynamics (CFD) code HARVEY [1] on the Summit supercomputer. We demonstrate minimal-overhead, in situ visualization relative to simulation alone, and compare the Halo Replay performance to tightly-coupled in situ approaches.
We evaluate Julia as a single language and ecosystem paradigm powered by LLVM to develop workflow components for high-performance computing. We run a Gray-Scott, 2-variable diffusion-reaction application using a memory-bound, 7-point stencil kernel on Frontier, the US Department of Energy’s first exascale supercomputer. We evaluate the performance, scaling, and trade-offs of (i) the computational kernel on AMD’s MI250x GPUs, (ii) weak scaling up to 4,096 MPI processes/GPUs or 512 nodes, (iii) parallel I/O writes using the ADIOS2 library bindings, and (iv) Jupyter Notebooks for interactive analysis. Results suggest that although Julia generates a reasonable LLVM-IR, a nearly 50% performance difference exists vs. native AMD HIP stencil codes when running on the GPUs. As expected, we observed near-zero overhead when using MPI and parallel I/O bindings for system-wide installed implementations. Consequently, Julia emerges as a compelling high-performance and high-productivity workflow composition language, as measured on the fastest supercomputer in the world.
Algorithmic differentiation (AD) is a tool for generating discrete adjoint solvers, which efficiently compute gradients of functions with many inputs, for example for use in gradient-based optimization. AD is often applied to large computations such as stencil operators, which are an important part of most structured-mesh PDE solvers. Stencil computations are often parallelized, for example by using OpenMP, and optimized by using techniques such as cache-blocking and tiling to fully utilize multicore CPUs and many-core accelerators and GPUs. Differentiating these codes with conventional reverse-mode AD results in adjoint codes that cannot be expressed as stencil operations and may not be easily parallelizable. They thus leave most of the compute power of modern architectures unused. We present a method that combines forward-mode AD and loop transformation to generate adjoint solvers that use the same memory access pattern as the original computation that they are derived from and can benefit from the same optimization techniques. The effectiveness of this method is demonstrated by generating a scalable adjoint CFD solver for multicore CPUs and Xeon Phi accelerators.
Stencil computations are not well optimized by general-purpose production compilers and the increased use of multicore, manycore, and accelerator-based systems makes the optimization problem even more challenging. In this paper we present Snowflake, a Domain Specific Language (DSL) for stencils that uses a 'micro-compiler' approach, i.e., small, focused, domain-specific code generators. The approach is similar to that used in image processing stencils, but Snowflake handles the much more complex stencils that arise in scientific computing, including complex boundary conditions, higher-order operators (larger stencils), higher dimensions, variable coefficients, non-unit-stride iteration spaces, and multiple input or output meshes. Snowflake is embedded in the Python language, allowing it to interoperate with popular scientific tools like SciPy and iPython; it also takes advantage of built-in Python libraries for powerful dependence analysis as part of a just-in-time compiler. We demonstrate the power of the Snowflake language and the micro-compiler approach with a complex scientific benchmark, HPGMG, that exercises the generality of stencil support in Snowflake. By generating OpenMP comparable to, and OpenCL within a factor of 2x of hand-optimized HPGMG, Snowflake demonstrates that a micro-compiler can support diverse processor architectures and is performance-competitive whilst preserving a high-level Python implementation.
We evaluate AI-assisted generative capabilities on fundamental numerical kernels in high-performance computing (HPC), including AXPY, GEMV, GEMM, SpMV, Jacobi Stencil, and CG. We test the generated kernel codes for a variety of language-supported programming models, including (1) C++ (e.g., OpenMP [including offload], OpenACC, Kokkos, SyCL, CUDA, and HIP), (2) Fortran (e.g., OpenMP [including offload] and OpenACC), (3) Python (e.g., numpy, Numba, cuPy, and pyCUDA), and (4) Julia (e.g., Threads, CUDA.jl, AMDGPU.jl, and KernelAbstractions.jl). We use the GitHub Copilot capabilities powered by the GPT-based OpenAI Codex available in Visual Studio Code as of April 2023 to generate a vast amount of implementations given simple + + prompt variants. To quantify and compare the results, we propose a proficiency metric around the initial 10 suggestions given for each prompt. Results suggest that the OpenAI Codex outputs for C++ correlate with the adoption and maturity of programming models. For example, OpenMP and CUDA score really high, whereas HIP is still lacking. We found that prompts from either a targeted language such as Fortran or the more general purpose Python can benefit from adding code keywords, while Julia prompts perform acceptably well for its mature programming models (e.g., Threads and CUDA.jl). We expect for these benchmarks to provide a point of reference for each programming model's community. Overall, understanding the convergence of large language models, AI, and HPC is crucial due to its rapidly evolving nature and how it is redefining human-computer interactions.
The SOLPS-ITER code suite is used worldwide for plasma edge modeling, the interpretation of experiments, as well as for the design of the ITER divertor. The numerical scheme of the plasma solver of the code, B2.5, is based on the assumption of perfectly field-aligned grids, while in practice grids are often strongly distorted to match divertor target shapes. Neglecting these grid distortion leads to qualitatively and quantitatively incorrect results for fluid neutral simulations, and may affect results in cold (detached) divertors even when using kinetic neutral simulations. In this contribution, we present the first results of a newly implemented 9-point stencil in B2.5 to properly handle misaligned grids. The new scheme is then applied to fluid neutral simulations of a well-diagnosed and previously modeled Alcator C-Mod discharge. Results are compared with the original 5-point scheme neglecting grid distortion effects, as well as with simulations including a full kinetic neutral model. We conclude that the 9-point stencil is essential to correctly model the transport of fluid neutrals on distorted grids, and to capture the effects of divertor closure on the fluid neutral behavior.
We apply AI-assisted large language model (LLM) capabilities of GPT-3 targeting high-performance computing (HPC) kernels for (i) code generation, and (ii) auto-parallelization of serial code in C ++, Fortran, Python and Julia. Our scope includes the following fundamental numerical kernels: AXPY, GEMV, GEMM, SpMV, Jacobi Stencil, and CG, and language/programming models: (1) C++ (e.g., OpenMP [including offload], OpenACC, Kokkos, SyCL, CUDA, and HIP), (2) Fortran (e.g., OpenMP [including offload] and OpenACC), (3) Python (e.g., numpy, Numba, cuPy, and pyCUDA), and (4) Julia (e.g., Threads, CUDA.jl, AMDGPU.jl, and KernelAbstractions.jl). Kernel implementations are generated using GitHub Copilot capabilities powered by the GPT-based OpenAI Codex available in Visual Studio Code given simple + + prompt variants. To quantify and compare the generated results, we propose a proficiency metric around the initial 10 suggestions given for each prompt. For auto-parallelization, we use ChatGPT interactively giving simple prompts as in a dialogue with another human including simple “prompt engineering” follow ups. Results suggest that correct outputs for C++ correlate with the adoption and maturity of programming models. For example, OpenMP and CUDA score really high, whereas HIP is still lacking. We found that prompts from either a targeted language such as Fortran or the more general-purpose Python can benefit from adding language keywords, while Julia prompts perform acceptably well for its Threads and CUDA.jl programming models. Finally, we expect to provide an initial quantifiable point of reference for code generation in each programming model using a state-of-the-art LLM. Overall, understanding the convergence of LLMs, AI, and HPC is crucial due to its rapidly evolving nature and how it is redefining human-computer interactions.
The Message Passing Interface (MPI) standard plays a crucial role in enabling scientific applications for parallel computing and is an essential component in high-performance computing (HPC). However, implementing MPI code manually—especially applying a proper domain decomposition and communication pattern—is a challenging and error-prone task. We present ChatMPI, an AI assistant for MPI parallelization of sequential C codes. In our analysis, we focus on testing six essential HPC workloads, which are based on Basic Linear Algebra Subprograms levels 1, 2, and 3 as well as sparse, stencil, and iterative operations. We analyze the process of creating ChatMPI by using the ChatHPC library. This lightweight large language model (LLM)–based infrastructure enables HPC experts to efficiently create and supervise trustworthy AI capabilities for critical HPC software tasks. We study the data required for training (fine-tuning) ChatMPI to generate parallel codes that not only use MPI syntax correctly but also apply HPC techniques to reduce memory communication and maximize performance by using proper work decomposition. With a relatively small training dataset composed of a few dozen prompts and fewer than 15 minutes of fine-tuning on one node equipped with two NVIDIA H100 GPUs, ChatMPI elevates trustworthiness for MPI code generation of current LLMs (e.g., Code Llama, ChatGPT-4o and ChatGPT 5). Additionally, we evaluate the performance of the MPI codes generated by ChatMPI in comparison with the ones generated by ChatGPT-4o and ChatGPT-5. The codes generated by ChatMPI provide up to a 4 × boost in performance by using better problem decomposition, communication patterns, and HPC techniques (e.g., communication avoiding).
In order to achieve exascale systems, application resilience needs to be addressed. Some programming models, such as task-DAG (directed acyclic graphs) architectures, currently embed resilience features whereas traditional SPMD (single program, multiple data) and message-passing models do not. Since a large part of the community's code base follows the latter models, it is still required to take advantage of application characteristics to minimize the overheads of fault tolerance. To that end, this paper explores how recovering from hard process/node failures in a local manner is a natural approach for certain applications to obtain resilience at lower costs in faulty environments. In particular, this paper targets enabling online, semitransparent local recovery for stencil computations on current leadership-class systems as well as presents programming support and scalable runtime mechanisms. Also described and demonstrated in this paper is the effect of failure masking, which allows the effective reduction of impact on total time to solution due to multiple failures. Furthermore, we discuss, implement, and evaluate ghost region expansion and cell-to-rank remapping to increase the probability of failure masking. To conclude, this paper shows the integration of all aforementioned mechanisms with the S3D combustion simulation through an experimental demonstration (using the Titan system) of the ability to tolerate high failure rates (i.e., node failures every five seconds) with low overhead while sustaining performance at large scales. In addition, this demonstration also displays the failure masking probability increase resulting from the combination of both ghost region expansion and cell-to-rank remapping.
FFTX-IRIS is a dynamic system to efficiently utilize novel heterogeneous platforms. This system links two next-generation frameworks, FFTX and IRIS, to navigate the complexity of different hardware architectures. FFTX provides a runtime code generation framework for high-performance Fast Fourier Transform kernels. IRIS runtime provides portability and multi-device heterogeneity, allowing computation on any available compute resource. Together, FFTX-IRIS enables code generation, seamless portability, and performance without user involvement. We show the design of the FFTX-IRIS system along with an evaluation of various small FFT benchmarks. We also demonstrate multi-device heterogeneity of FFTX-IRIS with a larger stencil application.
The need to scale next-generation industrial engineering problems to the largest computational platforms presents unique challenges. Such problems may have complex coupled multi-physics models, as well as computationally and communication intensive algorithms not often seen in standard academic problems, while also needing high-demand I/O for analysis purposes. In such cases even codes with good existing scaling properties may need significant changes, addressed in a cross-cutting way to solve such problems at extreme scales. These challenges relate to system components that were not previously problematic on less complex problems and/or at smaller computational scales. This paper illustrates these challenges for Uintah, a highly scalable asynchronous many-task runtime system being applied to the modeling of a 1000 MWe ultra-supercritical coal boiler. In order to model this formidable, production engineering problem, not only was an integrated approach needed that built upon existing scalable components in areas such as complex stencil calculations and linear algebra, but the significant challenges of high demand I/O, complex task graphs, and infrastructure inefficiencies also had to be addressed. This integrated approach allowed this problem to run on 256K CPU cores on Mira, with good weak scaling to 512K CPU cores and similar strong scaling, except for radiation. The need for strong scaling of a new ray tracing-based radiation model required substantial Uintah infrastructure improvements before 119K CPU cores and 7.5K GPUs on Titan could be used, with scaling out to 256K CPU cores and 16K GPUs. The resulting code demonstrates not only excellent overall scalability but a significant improvement in overall performance for the full-scale, production boiler problem.
Automated techniques for rigorous floating-point round-off error analysis are a prerequisite to placing important activities in HPC such as precision allocation, verification and code optimization on a formal footing. Yet existing techniques cannot provide tight bounds for expressions beyond a few dozen operators; barely enough for HPC. In this work, we offer an approach embedded in a new tool called SATIRE that scales error analysis by four orders of magnitude compared to today’s best-of-class tools. We explain how three key ideas underlying SATIRE help it attain such scale: path strength reduction, bound optimization and abstraction. SATIRE provides tight bounds and rigorous guarantees on significantly larger expressions with well over a hundred thousand operators, covering important examples including FFT, matrix multiplication and PDE stencils.
Iterative memory-bound solvers commonly occur in HPC codes. Typical GPU implementations have a loop on the host side that invokes the GPU kernel as much as time/algorithm steps there are. The termination of each kernel implicitly acts the barrier required after advancing the solution every time step. We propose an execution model for running memory-bound iterative GPU kernels: PERsistent KernelS (PERKS). In this model, the time loop is moved inside persistent kernel, and device-wide barriers are used for synchronization. We then reduce the traffic to device memory by caching subset of the output in each time step in the unused registers and shared memory. PERKS can be generalized to any iterative solver: they largely independent of the solver's implementation. We explain the design principle of PERKS and demonstrate effectiveness of PERKS for a wide range of iterative 2D/3D stencil benchmarks (geomean speedup of 2.12x for 2D stencils and 1.24x for 3D stencils over state-of-art libraries), and a Krylov subspace conjugate gradient solver (geomean speedup of 4.86x in smaller SpMV datasets from SuiteSparse and 1.43x in larger SpMV datasets over a state-of-art library). All PERKS-based implementations available at: https://github.com/neozhang307/PERKS.
Wind applications require the ability to simulate rotating blades. To support this use-case, a novel design-order sliding mesh algorithm has been developed and deployed. The hybrid method combines the control volume finite element methodology (CVFEM) with concepts found within a discontinuous Galerkin (DG) finite element method (FEM) to manage a sliding mesh. The method has been demonstrated to be design-order for the tested polynomial basis (P=1 and P=2) and has been deployed to provide production simulation capability for a Vestas V27 (225 kW) wind turbine. Other stationary and canonical rotating ow simulations are also presented. As the majority of wind-energy applications are driving extensive usage of hybrid meshes, a foundational study that outlines near-wall numerical behavior for a variety of element topologies is presented. Results indicate that the proposed nonlinear stabilization operator (NSO) is an effective stabilization methodology to control Gibbs phenomena at large cell Peclet numbers. The study also provides practical mesh resolution guidelines for future analysis efforts. Application-driven performance and algorithmic improvements have been carried out to increase robustness of the scheme on hybrid production wind energy meshes. Specifically, the Kokkos-based Nalu Kernel construct outlined in the FY17/Q4 ExaWind milestone has been transitioned to the hybrid mesh regime. This code base is exercised within a full V27 production run. Simulation timings for parallel search and custom ghosting are presented. As the low-Mach application space requires implicit matrix solves, the cost of matrix reinitialization has been evaluated on a variety of production meshes. Results indicate that at low element counts, i.e., fewer than 100 million elements, matrix graph initialization and preconditioner setup times are small. However, as mesh sizes increase, e.g., 500 million elements, simulation time associated with \setup-up" costs can increase to nearly 50% of overall simulation time when using the full Tpetra solver stack and nearly 35% when using a mixed Tpetra- Hypre-based solver stack. The report also highlights the project achievement of surpassing the 1 billion element mesh scale for a production V27 hybrid mesh. A detailed timing breakdown is presented that again suggests work to be done in the setup events associated with the linear system. In order to mitigate these initialization costs, several application paths have been explored, all of which are designed to reduce the frequency of matrix reinitialization. Methods such as removing Jacobian entries on the dynamic matrix columns (in concert with increased inner equation iterations), and lagging of Jacobian entries have reduced setup times at the cost of numerical stability. Artificially increasing, or bloating, the matrix stencil to ensure that full Jacobians are included is developed with results suggesting that this methodology is useful in decreasing reinitialization events without loss of matrix contributions. With the above foundational advances in computational capability, the project is well positioned to begin scientific inquiry on a variety of wind-farm physics such as turbine/turbine wake interactions.
Wind applications require the ability to simulate rotating blades. To support this use-case, a novel design-order sliding mesh algorithm has been developed and deployed. The hybrid method combines the control volume finite element methodology (CVFEM) with concepts found within a discontinuous Galerkin (DG) finite element method (FEM) to manage a sliding mesh. The method has been demonstrated to be design-order for the tested polynomial basis (P=1 and P=2) and has been deployed to provide production simulation capability for a Vestas V27 (225 kW) wind turbine. Other stationary and canonical rotating flow simulations are also presented. As the majority of wind-energy applications are driving extensive usage of hybrid meshes, a foundational study that outlines near-wall numerical behavior for a variety of element topologies is presented. Results indicate that the proposed nonlinear stabilization operator (NSO) is an effective stabilization methodology to control Gibbs phenomena at large cell Peclet numbers. The study also provides practical mesh resolution guidelines for future analysis efforts. Application-driven performance and algorithmic improvements have been carried out to increase robustness of the scheme on hybrid production wind energy meshes. Specifically, the Kokkos-based Nalu Kernel construct outlined in the FY17/Q4 ExaWind milestone has been transitioned to the hybrid mesh regime. This code base is exercised within a full V27 production run. Simulation timings for parallel search and custom ghosting are presented. As the low-Mach application space requires implicit matrix solves, the cost of matrix reinitialization has been evaluated on a variety of production meshes. Results indicate that at low element counts, i.e., fewer than 100 million elements, matrix graph initialization and preconditioner setup times are small. However, as mesh sizes increase, e.g., 500 million elements, simulation time associated with "setup-up" costs can increase to nearly 50% of overall simulation time when using the full Tpetra solver stack and nearly 35% when using a mixed Tpetra- Hypre-based solver stack. The report also highlights the project achievement of surpassing the 1 billion element mesh scale for a production V27 hybrid mesh. A detailed timing breakdown is presented that again suggests work to be done in the setup events associated with the linear system. In order to mitigate these initialization costs, several application paths have been explored, all of which are designed to reduce the frequency of matrix reinitialization. Methods such as removing Jacobian entries on the dynamic matrix columns (in concert with increased inner equation iterations), and lagging of Jacobian entries have reduced setup times at the cost of numerical stability. Artificially increasing, or bloating, the matrix stencil to ensure that full Jacobians are included is developed with results suggesting that this methodology is useful in decreasing reinitialization events without loss of matrix contributions. With the above foundational advances in computational capability, the project is well positioned to begin scientific inquiry on a variety of wind-farm physics such as turbine/turbine wake interactions.