Search NASA⌕ Search

SEARCH · Search NASA

Results for “Data parallelism”

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 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2022

Towards Low-Overhead Resilience for Data Parallel Deep Learning

Data parallel techniques have been widely adopted both in academia and industry as a tool to enable scalable training of deep learning models. At scale, DL training jobs can fail due to software or hardware bugs, may need to be preempted or terminated due to unexpected events, or may perform suboptimally because they were misconfigured. Under such circumstances, there is a need to recover and/or reconfigure data-parallel DL training jobs on-the-fly, while minimizing the impact on the accuracy of the DNN model and the runtime overhead. In this regard, state-of-art techniques adopted by the HPC community mostly rely on checkpoint-restart, which inevitably leads to loss of progress, thus increasing the runtime overhead. In this paper we explore alternative techniques that exploit the properties of modern deep learning frameworks (overlapping of gradient averaging and weight updates with local gradient computations through pipeline parallelism) to reduce the overhead of resilience/elasticity. To this end we introduce a failure simulation framework and two resilience strategies (immediate mini-batch rollback and lossy forward recovery), which we study compared with checkpoint-restart approaches in a variety of settings in order to understand the trade-offs between the accuracy loss of the DNN model and the runtime overhead.

data-parallel training↗
DOE OSTI2022

Accelerating Collective Communication in Data Parallel Training across Deep Learning Frameworks

This work develops new techniques within Horovod, a generic communication library supporting data parallel training across deep learning frameworks. In particular, we improve the Horovod control plane by implementing a new coordination scheme that takes advantage of the characteristics of the typical data parallel training paradigm, namely the repeated execution of collectives on the gradients of a fixed set of tensors. Using a caching strategy, we execute Horovod’s existing coordinator-worker logic only once during a typical training run, replacing it with a more efficient decentralized orchestration strategy using the cached data and a global intersection of a bitvector for the remaining training duration. Next, we introduce a feature for end users to explicitly group collective operations, enabling finer grained control over the communication buffer sizes. To evaluate our proposed strategies, we conduct experiments on a world-class supercomputer — Summit. We compare our proposals to Horovod’s original design and observe 2x performance improvement at a scale of 6000 GPUs; we also compare them against tf.distribute and torch.DDP and achieve 12% better and comparable performance, respectively, using up to 1536 GPUs; we compare our solution against BytePS in typical HPC settings and achieve about 20% better performance on a scale of 768 GPUs. Finally, we test our strategies on a scientific application (STEMDL) using up to 27,600 GPUs (the entire Summit) and show that we achieve a near-linear scaling of 0.93 with a sustained performance of 1.54 exaflops (with standard error +- 0.02) in FP16 precision.

Romero, Joshua↗
DOE OSTI2026

Memory-Aware External Facelist Calculation: A Data-Parallel Atomic Hash Counting Approach

Unstructured volumetric meshes serve as fundamental data representations in various scientific simulations and analyses. They play a crucial role in representing complex computational domains and are essential for important numerical techniques, such as finite element analysis. Whenever such a mesh is read from a file, streamed in-situ, or generated by algorithms, scientific visualization libraries rely on calculating the external surface of a geometry, named “external facelist”, to produce a polygonal mesh for rendering. Consequently, external facelist calculation has become one of the most widely used algorithms in the scientific visualization domain, necessitating optimal performance. In this paper, we explore relevant work on external facelist calculation algorithms in two common visualization libraries, VTK and Viskores, assess their performance and memory constraints, and introduce a novel memory-aware external facelist calculation algorithm employing an atomic hash counting approach. This algorithm fully leverages Viskores' data-parallel primitive operations, facilitating its execution across diverse many-core architectures. Our algorithm features the lowest memory footprint on the GPU and the second-lowest on the CPU among all evaluated methods, and it also delivers the fastest performance on both CPU and GPU. It has been made available under an open-source license in the VTK and Viskores visualization systems.

Tsalikis, Spiros [Kitware] (ORCID:0000000151137195↗
DOE OSTI2021

Optimization and Augmentation for Data Parallel Contour Trees

Contour trees are used for topological data analysis in scientific visualization. While originally computed with serial algorithms, recent work has introduced a vector-parallel algorithm. Furthermore, this algorithm is relatively slow for fully augmented contour trees which are needed for many practical data analysis tasks. We therefore introduce a representation called the hyperstructure that enables efficient searches through the contour tree and use it to construct a fully augmented contour tree in data parallel, with performance on average 6 times faster than the state-of-the-art parallel algorithm in the TTK topological toolkit.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

Rasterization with Data-Parallel Primitives

Parallel rasterization can suffer from race conditions during fragment generation, which is traditionally addressed by using specialized hardware accessible via vendor graphics APIs. Unfortunately, graphics APIs are increasingly problematic on high-performance computers, either because they are not provided or because of concerns about dependencies with in situ visualization. In response, we present a hardware-agnostic rasterization algorithm that handles race conditions using only data-parallel primitives (DPPs), enabling efficient rendering on HPC systems without graphics API dependencies and aligning with recent efforts to deliver visualization software with DPPs. Our evaluation consists of three phases: (1) evaluating portability across different CPU and GPU architectures, (2) evaluating competitiveness with a community standard, and (3) evaluating performance across varying workloads and available parallelism. The supporting experiments run on both AMD and NVIDIA GPUs, considering data sets as large as 460 million triangles and 160 million pixels. While performance generally falls short of graphics API baselines, it achieves interactive frame rates on most workloads. As a result, we conclude our approach is a viable solution for rasterization on high-performance computers since our approach is portably performant across different architectures without the need for specialized vendor support.

Buckley, Makani [University of Oregon] (ORCID:0009↗
DOE OSTI2020

Performance Analysis of Traditional and Data-Parallel Primitive Implementations of Visualization and Analysis Kernels

Measurements of absolute runtime are useful as a summary of performance when studying parallel visualization and analysis methods on computational platforms of increasing concurrency and complexity. We can obtain even more insights by measuring and examining more detailed measures from hardware performance counters, such as the number of instructions executed by an algorithm implemented in a particular way, the amount of data moved to/from memory, memory hierarchy utilization levels via cache hit/miss ratios, and so forth. This work focuses on performance analysis on modern multi-core platforms of three different visualization and analysis kernels that are implemented in different ways: one is "traditional", using combinations of C++ and VTK, and the other uses a data-parallel approach using VTK-m. Our performance study consists of measurement and reporting of several different hardware performance counters on two different multi-core CPU platforms. The results reveal interesting performance differences between these two different approaches for implementing these kernels, results that would not be apparent using runtime as the only metric.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

HyLiPoD: Parallel Particle Advection via a Hybrid of Lifeline Scheduling and Parallelization-Over-Data

Performance characteristics of parallel particle advection algorithms can vary greatly based on workload.With this short paper, we build a new algorithm based on results from a previous bake-off study which evaluated the performance of four algorithms on a variety of workloads. Our algorithm, called HyLiPoD, is a ''meta-algorithm,'' i.e., it considers the desired workload to choose from existing algorithms to maximize performance. To demonstrate HyliPoD's benefit, we analyze results from 162 tests including concurrencies of up to 8192 cores, meshes as large as 34 billion cells, and particle counts as large as 300 million. Our findings demonstrate that HyLiPoD's adaptive approach allows it to match the best performance of existing algorithms across diverse workloads.

high performance computing↗
DOE OSTI2021

ARENA: Asynchronous Reconfigurable Accelerator Ring to Enable Data-Centric Parallel Computing

The next generation HPC and data centers are likely to be reconfigurable and data-centric due to the trend of hardware specialization and the emergence of data-driven applications. In this work, we propose ARENA – an asynchronous reconfigurable accelerator ring architecture as a potential scenario on how the future HPC and data centers will be like. Despite using the coarse-grained reconfigurable arrays (CGRAs) as the substrate platform, our key contribution is not only the CGRA-cluster design itself, but also the ensemble of a new architecture and programming model that enables asynchronous tasking across a cluster of reconfigurable nodes, so as to bring specialized computation to the data rather than the reverse. We presume distributed data storage without asserting any prior knowledge on the data distribution. Hardware specialization occurs at runtime when a task finds the majority of data it requires are available at the present node. In other words, we dynamically generate specialized CGRA accelerators where the data reside. The asynchronous tasking for bringing computation to data is achieved by circulating the task token, which describes the dataflow graphs to be executed for a task, among the CGRA cluster connected by a fast ring network. Evaluations on a set of HPC and data-driven applications across different domains show that ARENA can provide better parallel scalability with reduced data movement (53.9 percent). Compared with contemporary compute-centric parallel models, ARENA can bring on average 4.37× speedup. The synthesized CGRAs and their task-dispatchers only occupy 2.93mm 2 chip area under 45nm process technology and can run at 800MHz with on average 759.8mW power consumption. ARENA also supports the concurrent execution of multi-applications, offering ideal architectural support for future high-performance parallel computing and data analytics systems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Optimizing Distributed Training on Frontier for Large Language Models

Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Loss scaling studies have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of 38.38%, 36.14%, and 31.96%, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved 100% weak scaling efficiency on 1024 and 3072 Mi250X GPUs, respectively. We also achieved strong scaling efficiencies of 89% and 87% for these two models. We trained these models only tens of iterations instead of training till completion.

Yin, Junqi↗
DOE OSTI2026

User Manual - HydraGNN v5.0: Distributed Implementation of Multi-Tasking Graph Neural Networks

This document serves as the user manual for HydraGNN v5.0, a scalable graph neural network (GNN) architecture for simultaneous prediction of multiple target properties using multi-task learning (MTL). This version of HydraGNN has been developed primarily to support the development, training, and deployment of predictive graph-based deep learning (DL) models for atomistic materials modeling. HydraGNN is templated over 13 message-passing policies, including invariant models (GIN, PNA, PNAPlus, GAT, MFC, CGCNN, SAGE, SchNet, DimeNet) and equivariant models (EGNN, PNAEq, PAINN, MACE), and supports distributed training via distributed data parallelism (DDP), DeepSpeed, and Fully Sharded Data Parallelism (FSDP) on leadership-class supercomputers. Although HydraGNN can be applied to problems beyond atomistic materials modeling, its current use is confined to homogeneous graphs. Additional capabilities include machine-learned interatomic potentials with energy-conserving forces, General, Powerful, and Scalable Graph Transformer (GraphGPS) global attention, periodic boundary conditions, hyperparameter optimization, mixed-precision training, and uncertainty quantification.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Pretraining Billion-Scale Geospatial Foundational Models on Frontier

As AI workloads increase in scope, generalization capability becomes challenging for small task-specific models and their demand for large amounts of labeled training samples increases. On the contrary, Foundation Models (FMs) are trained with internet-scale unlabeled data via self-supervised learning and have been shown to adapt to various tasks with minimal fine-tuning. Although large FMs have demonstrated significant impact in natural language processing and computer vision, efforts toward FMs for geospatial applications have been restricted to smaller size models, as pretraining larger models requires very large computing resources equipped with state-of-the-art hardware accelerators. Current satellite constellations collect 100+TBs of data a day, resulting in images that are billions of pixels and multimodal in nature. Such geospatial data poses unique challenges opening up new opportunities to develop FMs. We investigate billion scale FMs and HPC training profiles for geospatial applications by pretraining on publicly available data. We studied from end-to-end the performance and impact in the solution by scaling the model size. Our larger 3B parameter size model achieves up to 30% improvement in top1 scene classification accuracy when comparing a 100M parameter model. Moreover, we detail performance experiments on the Frontier supercomputer, America's first exascale system, where we study different model and data parallel approaches using PyTorch's Fully Sharded Data Parallel library. Specifically, we study variants of the Vision Transformer architecture (ViT), conducting performance analysis for ViT models with size up to 15B parameters. By discussing throughput and performance bottlenecks under different parallelism configurations, we offer insights on how to leverage such leadership-class HPC resources when developing large models for geospatial imagery applications.

Tsaris, Aristeidis (aris)↗
DOE OSTI2021

Asynchronous and Load-Balanced Union-Find for Distributed and Parallel Scientific Data Visualization and Analysis

We present a novel distributed union-find algorithm that features asynchronous parallelism and k-d tree based load balancing for scalable visualization and analysis of scientific data. Applications of union-find include level set extraction and critical point tracking, but distributed union-find can suffer from high synchronization costs and imbalanced workloads across parallel processes. In this study, we prove that global synchronizations in existing distributed union-find can be eliminated without changing final results, allowing overlapped communications and computations for scalable processing. We also use a k-d tree decomposition to redistribute inputs, in order to improve workload balancing. We benchmark the scalability of our algorithm with up to 1,024 processes using both synthetic and application data. Here, we demonstrate the use of our algorithm in critical point tracking and super-level set extraction with high-speed imaging experiments and fusion plasma simulations, respectively.

97 MATHEMATICS AND COMPUTING↗