Search NASA⌕ Search

SEARCH · Search NASA

Results for “high performance computing systems”

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.

At least 91 records · Page 5
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

High-Resolution Simulations of Geological CO 2 Injection: Application to the SPE11 Benchmark

Geological carbon sequestration (GCS) will play a critical role in decarbonization and in facilitating the transition to clean energy systems. Because CO 2 is highly mobile, ensuring its safe and permanent injection into subsurface geological formations involves monitoring over larger spatial domains and longer time periods than is typical for hydrocarbon reservoirs. This can benefit from simulation tools capable of modeling key CO 2 trapping mechanisms, particularly those optimized for speed and scalability on high-performance computing systems. Using isothermal versions of the SPE11B and SPE11C benchmark cases, we conduct a mesh refinement study simulating CO 2 injection into kilometer-scale rock formations at centimeter resolution with the GEOS open-source simulation framework. We focus on how mesh refinement improves the accuracy of convective mixing in both 2D and 3D simulations. The computational costs associated with achieving a converged solution highlight the need for predictive upscaling techniques. A systematic performance scaling analysis—including both central processing unit (CPU) and graphics processing unit (GPU) architectures—complements the “Results” section.

Geosciences↗
DOE OSTI2023

Computing the Properties of Matter with Leadership Computing Resources (Closeout Report for DE-SC0018121)

In order to add more capabilities to Halide, we have designed a new framework called Tiramisu and integrated this framework into Halide. Since Tiramisu enables Halide to target heterogeneous architectures, our development efforts have been refocused on Tiramisu. Most high-performance computer systems today are complex and increasingly heterogeneous; they may have CPUs, GPUs and FPGAs. Achieving best performance requires taking full advantage of all these different architectures. To address this issue, we have designed Tiramisu, an optimization framework that enables Halide (and other DSLs) to target heterogeneous architectures. Tiramisu is an optimization framework that takes as input a high level, architecture-independent representation of code and a set of scheduling and data mapping commands that guide code transformation. The input can either be generated by a domain-specific language (DSL) compiler such as Halide or directly written by a programmer. Tiramisu then applies the user-specified code and data-layout transformations and generates an architecture-specific, low-level intermediate representation (IR) that takes advantage of modern architectural features such as multicore parallelism, non-uniform memory (NUMA) hierarchies, clusters, and accelerators like GPUs and FPGAs. We integrated Tiramisu within Halide and implemented a representative set of benchmarks to evaluate this integration. Tiramisu is now open source and is available for public use (http://tiramisu-compiler.org/). A paper about Tiramisu was published, it shows that Tiramisu extends Halide with many new capabilities and that Tiramisu can generate efficient code for multicores, GPUs, FPGAs and distributed heterogeneous systems. The performance of code generated by the Tiramisu backends matches or exceeds hand optimized reference implementations. For example, the multicore backend matches the highly optimized Intel MKL library on many kernels and shows speedups reaching 4x over the original Halide. In addition to making Tiramisu more robust, we have used Tiramisu to implement a set of representative tensor operation for constructing baryon building blocks required for multi baryon contractions in LQCD. In order to implement this code, we needed to generalize Tiramisu in two ways: first we needed to support indirect array accesses, and second, we needed to add support for complex numbers to Tiramisu. The code generated by Tiramisu is 6x faster than the reference code. Our efforts towards an MPI based multi-node version of tiramisu have matured and the resulting code scales well on multiple nodes (tests up to 512 KNL nodes have been undertaken).

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2023

Neglecting Model Parametric Uncertainty Can Drastically Underestimate Flood Risks

Abstract Floods drive dynamic and deeply uncertain risks for people and infrastructures. Uncertainty characterization is a crucial step in improving the predictive understanding of multi‐sector dynamics and the design of risk‐management strategies. Current approaches to estimate flood hazards often sample only a relatively small subset of the known unknowns, for example, the uncertainties surrounding the model parameters. This approach neglects the impacts of key uncertainties on hazards and system dynamics. Here we mainstream a recently developed method for Bayesian inference to calibrate a computationally expensive distributed hydrologic model. We compare three different calibration approaches: (a) stepwise line search, (b) precalibration or screening, and (c) the Fast Model Calibrations (FaMoS) approach. FaMoS deploys a particle‐based approach that takes advantage of the massive parallelization afforded by modern high‐performance computing systems. We quantify how neglecting parametric uncertainty and data discrepancy can drastically underestimate extreme flood events and risks. Precalibration improves prediction skill score over a stepwise line search. The Bayesian calibration improves the uncertainty characterization of model parameters and flood risk projections.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

Extreme-scale EV charging infrastructure planning for last-mile delivery using high-performance parallel computing

Here, this paper addresses stochastic charger location and allocation problems under queue congestion for last-mile delivery using electric vehicles (EVs). The objective is to decide where to open charging stations and how many chargers of each type to install, subject to budgetary and waiting-time constraints. We formulate the problem as a mixed-integer non-linear program, where each station-charger pair is modeled as a multiserver queue with stochastic arrivals and service times to capture the notion of waiting in fleet operations. The model is extremely large, with billions of variables and constraints for a typical metropolitan area; even loading the model in solver memory is difficult, let alone solving it. To address this challenge, we develop a Lagrangian-based dual decomposition framework that decomposes the problem by station and leverages parallelization on high-performance computing systems, where the subproblems are solved by using a cutting plane method and their solutions are collected at the master level. We also develop a three-step rounding heuristic to transform the fractional subproblem solutions into feasible integral solutions. Computational experiments on data from the Chicago metropolitan area with hundreds of thousands of households and thousands of candidate stations show that our approach produces high-quality solutions in cases where existing exact methods cannot even load the model in memory. We also analyze various policy scenarios, demonstrating that combining existing depots with newly built stations under multiagency collaboration substantially reduces costs and congestion. These findings offer a scalable and efficient framework for developing sustainable large-scale EV charging networks.

Capacity allocation↗
DOE OSTI2017

The Trinity Center of Excellence Co-Design Best Practices

Co-design between application developers and the vendors providing high-performance computing systems is essential to efficiently using current-generation HPC systems and shaping next-generation architectures. Application developers are faced with the challenge of effectively orchestrating many hardware threads and exploiting vector instructions, which can require code modifications. Will refactoring applications for current-generation systems apply to future-generation systems? Is there help available for this enormous task? Over the past 10 years, a partnership has evolved to address these issues. Centers of Excellence are mutually beneficial collaborations between application developers and vendors; vendors receive a better understanding of the needs of their users, and users receive a better understanding of the characteristics of the hardware, system software, and tools available to address their challenges. Lastly, this article describes the partnership between two vendors, Cray and Intel, and the application teams working together in the Trinity Center of Excellence.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

CommBench: Micro-Benchmarking Hierarchical Networks with Multi-GPU, Multi-NIC Nodes

Modern high-performance computing systems have multiple GPUs and network interface cards (NICs) per node. The resulting network architectures have multilevel hierarchies of subnetworks with different interconnect and software technologies. These systems offer multiple vendor-provided communication capabilities and library implementations (IPC, MPI, NCCL, RCCL, OneCCL) with APIs providing varying levels of performance across the different levels. Understanding this performance is currently difficult because of the wide range of architectures and programming models (CUDA, HIP, OneAPI). We present CommBench, a library with cross-system portability and a high-level API that enables developers to easily build microbenchmarks relevant to their use cases and gain insight into the performance (bandwidth & latency) of multiple implementation libraries on different networks. We demonstrate CommBench with three sets of microbenchmarks that profile the performance of six systems. Our experimental results reveal the effect of multiple NICs on optimizing the bandwidth across nodes and also present the performance characteristics of four available communication libraries within and across nodes of NVIDIA, AMD, and Intel GPU networks.

Hidayetoglu, Mert↗
DOE OSTI2022

Computer Science Research Needs for Parallel Discrete Event Simulation (PDES)

Historically, scientific computing efforts have demonstrated the clear need for, and effective use of, supercomputing with traditional time-stepped simulations. Nevertheless, there are several areas in the mission spaces of the U.S. Department of Energy and other agencies waiting to tap advanced computing research using a different, discrete event style of modeling, simulation, and analysis. These span a wide spectrum of applications including energy grid resilience, urban planning and policy, transportation science, building technologies, emergency response and planning, environmental impact analysis, computational epidemiology, Internet communications, cyber security, and cyber-physical systems, to name only a few. Even within traditional scientific applications, the role of discrete event modes of execution is increasing in the form of new event-based mathematical solvers such as quantized state integration methods and discrete-continuous hybrid system solvers. Co-design of advanced supercomputing hardware systems is another area that exploits discrete event simulation at its core for effective analyses. Complex systems, entity behaviors and interconnections play a significant role in all these applications, which are mapped to large-scale models with discrete event formulations. To make advancements in all the aforementioned scientific areas, many technical aspects need to be more thoroughly studied and deeply understood in parallel discrete event simulation (PDES). The unique dynamics inherent in a discrete event modeling approach, by their very nature, intersect and influence the entire stack of the computing system, including (a) the unique nature of the instruction sets exercised in PDES workloads without a predominance of high-precision floating point operations, (b) virtual time-constrained multi-threaded execution of many logical processes per processor, (c) extremely variable and difficult to predict network traffic characteristics, (d) interfaces and inter-dependencies with machine learning and artificial intelligence codes at higher software layers, and (e) highly challenging load balancing needs, especially in effectively accounting for accelerated/extremely heterogeneous computing in current and future high-performance computing systems. Efficient and accurate parallel execution of PDES workloads is also dominated by challenges in dealing with their asynchronous concurrency fundamentally present at the model level. Conservative synchronization, optimistic/speculative synchronization, and their hybrid schemes open new questions in fundamental computer science with respect to reversibility of computation and prediction (lookahead) of behaviors inherent within model codes. On the implementation front, there are relatively few scalable, general-purpose parallel discrete event simulators in the world, and even fewer have been studied on emerging hardware platforms. To enable scientific advances using PDES, the research needs in computer science must also be pursued and met in the intersection of the algorithmic and hardware-aware aspects of scalable PDES engines. This report is aimed at capturing a computer science-oriented view of this important area of research in PDES, presenting a sample of important applications with their inherent discrete event technology elements. Needs are outlined in core areas of parallel discrete event research as well as cross-cutting directions in computer science research that positively impact scientific advancements across several important application areas. A selection of priority research opportunities in advanced computing for PDES is identified to serve as reference for key research topics and their order of importance for scientific advancements.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Parallelization and Performance Portability in Hydrodynamics Codes

With the eve of Exascale computing, performance and portability are at the forefront of all scientific codes. Adding more cores and more energy to a system is no longer a sustainable way to achieve performance, and extra effort must now be made to improve performance in all areas of code and code development. Using hydrodynamic codes as a basis, this work explores numerous techniques to achieve performance in different ways. Adaptive mesh refinement (AMR) is a necessary technique to improve memory optimization in mesh-based simulations. However it is invasive and conventionally difficult to integrate into existing applications, so we present a new branch of AMR to create a smooth transition to these optimizations, which not only improves performance, but also greatly reduces developer effort. We introduce the concept of this improvement as Phantom-Cell AMR, and assess theoretically the improvements, as well as present an application of its use. Other work included involves and investigation into an efficient data structure that ensures optimal memory layout for cache performance, with a target of making codes performant and portable across all architectures. All of the work targets both performance and portability, not just on CPU hardware, but specifically across GPU architectures. Parallel performance is key to all of the methods presented, but the research makes a great effort to improve the portability of all applications to prepare for current high performance computing systems and those on the horizon.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

A Scalable Real-Time Data Assimilation Framework for Predicting Turbulent Atmosphere Dynamics

AI-based foundation models like FourCastNet, GraphCast are revolutionizing weather and climate predictions but are not yet ready for operational use. Their limitation lies in the absence of a data assimilation system to incorporate real-time Earth system observations, crucial for accurately forecasting events like tropical cyclones. To overcome these obstacles, we introduce a generic real-time data assimilation framework and demonstrate its end-to-end performance on the Frontier supercomputer. This framework comprises two primary modules: an ensemble score filter (EnSF), which significantly outperforms the state-of-the-art data assimilation method, and a vision transformer-based surrogate capable of real-time adaptation through the integration of observational data. We demonstrate both the strong and weak scaling of our framework up to 1024 GPUs on the Exascale supercomputer, Frontier. Our results not only illustrate the framework's exceptional scalability on high-performance computing systems, but also demonstrate the importance of supercomputers in real-time data assimilation for weather and climate predictions.

Lu, Dan↗
DOE OSTI2025

FY25 MOOSE Usability Improvements: 3D Meshing Capabilities, Initiation of Geometry Support for Monte Carlo Tools, and Enhancement of MOOSE/Workbench User Input Interactions

Usability improvements have been made to MOOSE and Workbench in FY25 to enhance usability and user workflows. Assorted enhancement have been made to MOOSE’s intrinsic meshing capabilities in order to enable more flexible and complex meshing of nuclear reactor systems, in particular for 3D applications. Mesh generators have been added to perform operations such as batch mesh generation, surface mesh generation, and creation of 3D transition layers. These mesh generation capabilities make it much easier to generate high quality non-extruded 3D meshes. Additionally, work to integrate Monte Carlo reactor physics simulations into MOOSE-based multi-physics workflows has reached another milestone with the implementation of the Constructive Solid Geometry (CSG) base framework. This framework lays the foundation for mesh generators to offer the user a generic CSG output option (as opposed to a finite element mesh). To support users, workshop on the MOOSE Reactor Module was delivered which featured hands-on examples using the NEAMS Workbench on INL’s High Performance Computing system. Recent updates to the NEAMS Workbench, WASP, and the MOOSE language server have introduced several improvements aimed at making MOOSE-based simulation setup and input management faster, more accurate, and easier to use. Key capabilities that have been added include multi-tab-stop autocompletion, visual input diagnostics, developer-directed data visualizations, upgraded ParaView integration, and Workspace-level file tracking. Together, these changes make it easier for users to build, validate, and manage complex MOOSE-based simulation models — especially those involving reusable components, included files, and datasets. The improvements are designed to save time, reduce input errors, and help users get to a successful simulation run faster, with more confidence in the results.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2023

Technical descriptions of the experimental dynamical downscaling simulations over North America by the CAM–MPAS variable-resolution model

Abstract. Comprehensive assessment of climate datasets is important for communicating model projections and associated uncertainties to stakeholders. Uncertainties can arise not only from assumptions and biases within the model but also from external factors such as computational constraint and data processing. To understand sources of uncertainties in global variable-resolution (VR) dynamical downscaling, we produced a regional climate dataset using the Model for Prediction Across Scales (MPAS; dynamical core version 4.0) coupled to the Community Atmosphere Model (CAM; version 5.4), which we refer to as CAM–MPAS hereafter. This document provides technical details of the model configuration, simulations, computational requirements, post-processing, and data archive of the experimental CAM–MPAS downscaling data. The CAM–MPAS model is configured with VR meshes featuring higher resolutions over North America as well as quasi-uniform-resolution meshes across the globe. The dataset includes multiple uniform- (240 and 120 km) and variable-resolution (50–200, 25–100, and 12–46 km) simulations for both the present-day (1990–2010) and future (2080–2100) periods, closely following the protocol of the North American Coordinated Regional Climate Downscaling Experiment. A deviation from the protocol is the pseudo-warming experiment for the future period, using the ocean boundary conditions produced by adding the sea surface temperature and sea-ice changes from the low-resolution version of the Max Planck Institute Earth System Model (MPI-ESM-LR) in the Coupled Model Intercomparison Project Phase 5 to the present-day ocean state from a reanalysis product. Some unique aspects of global VR models are evaluated to provide background knowledge to data users and to explore good practices for modelers who use VR models for regional downscaling. In the coarse-resolution domain, strong resolution sensitivity of the hydrological cycles exists over the tropics but does not appear to affect the midlatitude circulations in the Northern Hemisphere, including the downscaling target of North America. The pseudo-warming experiment leads to similar responses of large-scale circulations to the imposed radiative and boundary forcings in the CAM–MPAS and MPI-ESM-LR models, but their climatological states in the historical period differ over various regions, including North America. Such differences are carried to the future period, suggesting the importance of the base state climatology. Within the refined domain, precipitation statistics improve with higher resolutions, and such statistical inference is verified to be negligibly influenced by horizontal remapping during post-processing. Limited (≈50 % slower) throughput of the current code is found on a recent many-core/wide-vector high-performance computing system, which limits the lengths of the 12–46 km simulations and indirectly affects sampling uncertainty. Our experience shows that global and technical aspects of the VR downscaling framework require further investigations to reduce uncertainties for regional climate projection.

58 GEOSCIENCES↗
DOE OSTI2025

On the optimal sizing of power take-off systems for wave energy converters

The power take-off (PTO) system in a wave energy converter (WEC) is the means by which the energy in an ocean wave is converted into useful energy. There has been a general lack of technology convergence in the wave energy industry and PTO design has been largely device-specific with the design process determined by the WEC developer. An optimized PTO design can significantly affect the efficiency, reliability, performance, and overall cost and viability of a device. There remains a need in the marine energy industry for best practices and efficient design processes to be identified as it could greatly benefit developers and researchers in designing an optimal WEC PTO. This research utilizes the open-source Matlab-Simulink based software, WEC-Sim, which was developed by the National Renewable Energy Laboratory, and the Eagle, a high-performance computing system, to model and optimize two WECs of different archetypes and then investigates if any trends exist that can be exploited for greater efficiency in optimal PTO design. The results from this research indicate that there may be an optimal sizing for WEC PTOs for both power and PTO force rating and that substantial reductions can be made in the cost of a device without significant loss in the energy produced. Furthermore this optimal sizing may be independent of WEC type and deployment location. With an optimal, or near optimal, approach to WEC PTO sizing demonstrated, a methodology is proposed to address the challenge of nameplate ratings in the wave energy industry.

13 HYDRO ENERGY↗
DOE OSTI2019

PHISICS User Manual

PHISICS (Parallel and Highly Innovative Simulation for INL Code System) is a reactor physics package developed at the Idaho National Laboratory. It is composed of several modules: a nodal and semi-structured transport core solver (INSTANT), a depletion module (MRTAU), a time-dependent solver (TimeIntegrator), a cross-section interpolation and manipulation framework (MIXER), a SuperHomogenization scheme (SPH) a criticality search module (CRITICALITY), a fuel management and shuffling component (SHUFFLE) and a coupling API with RELAP5-3D. PHISICS has been built in a highly modular fashion so that different physical models can be activated or deactivated depending on the physics relevant to the particular type of analysis. This approach simplifies development as well as mid- and long-term maintenance of the code. PHISICS is entirely written in FORTRAN 95/2003. Each of its components is fully parallelized and designed to take full advantage of shared-memory computers (multiprocessor desktops and workstations) and middle to large high-performance computing systems (10 to 1000 processors). The aim of this document is to detail the input requirements for PHISICS (and all its modules) focusing on the input structure.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2024

Custom Accessors: Enabling Scalable Data Ingestion, (Re-)Organization, and Analysis on Distributed Systems

The emerging class of high velocity and high volume data analytic workflows comprise interwoven data ingestion, organization, and processing stages, with ingestion and organization steps often contributing comparable or even higher computational costs than actual processing steps. Since complex workflows consist of a variety of phases that view and use data differently, being able to construct efficient, scalable, distributed data structures (arrays, vectors, sets, maps, and multi-maps) is essential and requires custom methods to extend and shrink containers, analyze and position data, and, maintain globallyconsistent meta-data. In this paper, we propose a novel datastructure access paradigm based on the concept of Accessors. At a high level, accessors are customizable callable objects that can modify the behavior of insert, read, update, and delete operations for distributed containers while preserving atomicity guarantees. Accessors provide a very clean and natural way to implement a variety of programming patterns, e.g., conditional insertion/deletion and cascading computations, which would be otherwise hard (or even impossible) to express in parallel and distributed settings without using locks. We demonstrate the practicality and usefulness of our approach with two representative use cases and study the performance of these applications on a distributed High-Performance Computing system. Our analysis highlights that our proposed abstraction allows for an effective overlapping and concurrent execution of different workflow steps (e.g., data ingestion and analysis), which in a conventional analytics pipeline would execute sequentially, contributing cumulatively to the overall latency.

Castellana, Vito G. [BATTELLE (PACIFIC NW LAB)] (O↗
DOE OSTI2025

Power-Capping Metric Evaluation for Improving Energy Efficiency in HPC Applications

With high-performance computing systems now running at exascale, optimizing power-scaling management and resource utilization has become more critical than ever. This paper explores runtime power-capping optimizations that leverage integrated CPU-GPU power management on architectures like the NVIDIA GH200 superchip. We evaluate energy-performance metrics that account for simultaneous CPU and GPU power-capping effects by using two complementary approaches: speedup-energy-delay and a Euclidean distance-based multi-objective optimization method. By targeting a mostly compute-bound exascale science application, the Locally Self-Consistent Multiple Scattering (LSMS), we explore challenging scenarios to identify potential opportunities for energy savings in exascale applications, and we recognize that even modest reductions in energy consumption can have significant overall impacts. Our results highlight how GPU task-specific dynamic power-cap adjustments combined with integrated CPU-GPU power steering can improve the energy utilization of certain GPU tasks, thereby laying the groundwork for future adaptive optimization strategies.

Patrou, Maria [ORNL] (ORCID:0000000339754638)↗
DOE OSTI2019

Statistical and machine learning models for optimizing energy in parallel applications

Rising power costs and constraints are driving a growing focus on the energy efficiency of high performance computing systems. The unique characteristics of a particular system and workload and their effect on performance and energy efficiency are typically difficult for application users to assess and to control. Settings for optimum performance and energy efficiency can also diverge, so we need to identify trade-off options that guide a suitable balance between energy use and performance. We present statistical and machine learning models that only require a small number of runs to make accurate Pareto-optimal trade-off predictions using parameters that users can control. We study model training and validation using several parallel kernels and more complex workloads, including Algebraic Multigrid (AMG), Large-scale Atomic Molecular Massively Parallel Simulator, and Livermore Unstructured Lagrangian Explicit Shock Hydrodynamics. Here, we demonstrate that we can train the models using as few as 12 runs, with prediction error of less than 10%. Our AMG results identify trade-off options that provide up to 45% improvement in energy efficiency for around 10% performance loss. We reduce the sample measurement time required for AMG by 90%, from 13 h to 74 min.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

Performance and power modeling and prediction using MuMMI and 10 machine learning methods

Energy-efficient scientific applications require insight into how high performance computing system features impact the applications' power and performance. This insight can result from the development of performance and power models. Here, in this article, we use the modeling and prediction tool MuMMI (Multiple Metrics Modeling Infrastructure) and 10 machine learning methods to model and predict performance and power consumption and compare their prediction error rates. We use an algorithm-based fault-tolerant linear algebra code and a multilevel checkpointing fault-tolerant heat distribution code to conduct our modeling and prediction study on the Cray XC40 Theta and IBM BG/Q Mira at Argonne National Laboratory and the Intel Haswell cluster Shepard at Sandia National Laboratories. Our experimental results show that the prediction error rates in performance and power using MuMMI are less than 10% for most cases. By utilizing the models for runtime, node power, CPU power, and memory power, we identify the most significant performance counters for potential application optimizations, and we predict theoretical outcomes of the optimizations. Based on two collected datasets, we analyze and compare the prediction accuracy in performance and power consumption using MuMMI and 10 machine learning methods.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Distributed out-of-memory NMF on CPU/GPU architectures

We propose an efficient distributed out-of-memory implementation of the non-negative matrix factorization (NMF) algorithm for heterogeneous high-performance-computing systems. The proposed implementation is based on prior work on NMFk, which can perform automatic model selection and extract latent variables and patterns from data. In this work, we extend NMFk by adding support for dense and sparse matrix operation on multi-node, multi-GPU systems. The resulting algorithm is optimized for out-of-memory problems where the memory required to factorize a given matrix is greater than the available GPU memory. Memory complexity is reduced by batching/tiling strategies, and sparse and dense matrix operations are significantly accelerated with GPU cores (or tensor cores when available). Input/output latency associated with batch copies between host and device is hidden using CUDA streams to overlap data transfers and compute asynchronously, and latency associated with collective communications (both intra-node and inter-node) is reduced using optimized NVIDIA Collective Communication Library (NCCL) based communicators. Benchmark results show significant improvement, from 32X to 76x speedup, with the new implementation using GPUs over the CPU-based NMFk. Good weak scaling was demonstrated on up to 4096 multi-GPU cluster nodes with approximately 25,000 GPUs when decomposing a dense 340 Terabyte-size matrix and an 11 Exabyte-size sparse matrix of density 10 -6 .

97 MATHEMATICS AND COMPUTING↗