Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Distributed training”

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 OSTI2020

Data optimization for large batch distributed training of deep neural networks

Distributed training in deep learning (DL) is common practice as data and models grow. The current practice for distributed training of deep neural networks faces the challenges of communication bottlenecks when operating at scale, and model accuracy deterioration with an increase in global batch size. Present solutions focus on improving message exchange efficiency as well as implementing techniques to tweak batch sizes and models in the training process. The loss of training accuracy typically happens because the loss function gets trapped in a local minima. We observe that the loss landscape minimization is shaped by both the model and training data and propose a data optimization approach that utilizes machine learning to implicitly smooth out the loss landscape resulting in fewer local minima. Our approach filters out data points which are less important to feature learning, enabling us to speed up the training of models on larger batch sizes to improved accuracy.

Gahlot, Shubhankar↗
DOE OSTI2021

Distributed Training for High Resolution Images: A Domain and Spatial Decomposition Approach

In this work we developed two Pytorch libraries using the PyTorch RPC interface for distributed deep learning approaches on high resolution images. The spatial decomposition library allows for distributed training on very large images, which otherwise wouldn’t be possible on a single GPU. The domain parallelism library allows for distributed training across multiple domain unlabeled data, by leveraging the domain separation architecture. Both of those libraries where tested on the Summit supercomputer at Oak Ridge National Laboratory at a moderate scale.

Tsaris, Aristeidis (aris)↗
DOE OSTI2021

Elastic distributed training with fast convergence and efficient resource utilization

Distributed learning is now routinely conducted on cloud as well as dedicated clusters. Training with elastic resources brings new challenges and design choices. Prior studies focus on runtime performance and assume a static algorithmic behavior. In this work, by analyzing the impact of of resource scaling on convergence, we introduce schedules for synchronous stochastic gradient descent that proactively adapt the number of learners to reduce training time and improve convergence. Our approach no longer assumes a constant number of processors throughout training. In our experiment, distributed stochastic gradient descent with dynamic schedules and reduction momentum achieves better convergence and significant speedups over prior static ones. Numerous distributed training jobs running on cloud may benefit from our approach.

Cong, Guojing↗
DOE OSTI2021

Distributed Training for High Resolution Images: A Domain and Spatial Decomposition Approach

In this work we developed two Pytorch libraries using the PyTorch RPC interface for distributed deep learning approaches on high resolution images. The spatial decomposition library allows for distributedtraining on very large images, which otherwise won’t be possible on a single GPU. The domain parallelism library allows for distributed training across multiple domain unlabeled data, by leveraging the domain separation architecture. Both of those libraries were tested on the Summit supercomputer at a moderate scale, and we are releasing the code for both of them.

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 OSTI2022

Distributed Training and Optimization of Neural Networks

Deep learning models are yielding increasingly better performances thanks to multiple factors. To be successful, model may have large number of parameters or complex architectures and be trained on large dataset. This leads to large requirements on computing resource and turn around time, even more so when hyperparameter optimization is done (e.g. search over model architectures). While this is a challenge that goes beyond particle physics, we review the various ways to do the necessary computations in parallel, and put it in the context of high-energy physics.

Vlimant, Jean-Roch↗
DOE OSTI2022

Anderson Acceleration for Distributed Training of Deep Learning Models

Anderson acceleration (AA) is an extrapolation technique that has recently gained interest in the deep learning (DL) community to speed-up the sequential training of DL models. However, when performed at large scale, the DL training is exposed to a higher risk of getting trapped into steep local minima of the training loss function, and standard AA does not provide sufficient acceleration to escape from these steep local minima. This results in poor generalizability and makes AA ineffective. To restore AA’s advantage to speed-up the training of DL models on large scale computing platforms, we combine AA with an adaptive moving average procedure that boosts the training to escape from steep local minima. By monitoring the relative standard deviation between consecutive iterations, we also introduce a criterion to automatically assess whether the moving average is needed. We applied the method to the following DL instantiations for image classification: (i) ResNet50 trained on the open-source CIFAR100 dataset and (ii) ResNet50 trained on the open-source ImageNet1k dataset. Numerical results obtained using up to 1,536 NVIDIA V100 GPUs on the OLCF supercomputer Summit showed the stabilizing effect of the moving average on AA for all the problems above.

Lupo Pasini, Massimiliano↗
DOE OSTI2026

Tula: Optimizing Time, Cost, and Generalization in Distributed Large-Batch Training

Distributed training increases the number of batches processed per iteration either by scaling-out (adding more nodes) or scaling-up (increasing the batch-size). However, the largest configuration does not necessarily yield the best performance. Horizontal scaling introduces additional communication overhead, while vertical scaling is constrained by computation cost and device memory limits. Thus, simply increasing the batch-size leads to diminishing returns: training time and cost decrease initially but eventually plateaus, creating a knee-point in the time/cost vs. batch-size pareto curve. The optimal batch-size therefore depends on the underlying model, data and available compute resources. Large batches also suffer from worse model quality due to the well-known “generalization gap”. In this paper, we present Tula, an online service that automatically optimizes time, cost, and convergence quality for large-batch training of convolutional models. It combines parallel-systems modeling with statistical performance prediction to identify the optimal batchsize. Tula predicts training time and cost within 7.5−14% error across multiple models, and achieves up to 20× overall speedup and improves test accuracy by ≈9% on average over standard large-batch training on various vision tasks, thus successfully mitigating the generalization gap and accelerating training at the same time.

Tyagi, Sahil [ORNL] (ORCID:0009000783144745)↗
DOE OSTI2021

Federated Quantum Machine Learning

Distributed training across several quantum computers could significantly improve the training time and if we could share the learned model, not the data, it could potentially improve the data privacy as the training would happen where the data is located. One of the potential schemes to achieve this property is the federated learning (FL), which consists of several clients or local nodes learning on their own data and a central node to aggregate the models collected from those local nodes. However, to the best of our knowledge, no work has been done in quantum machine learning (QML) in federation setting yet. In this work, we present the federated training on hybrid quantum-classical machine learning models although our framework could be generalized to pure quantum machine learning model. Specifically, we consider the quantum neural network (QNN) coupled with classical pre-trained convolutional model. Our distributed federated learning scheme demonstrated almost the same level of trained model accuracies and yet significantly faster distributed training. It demonstrates a promising future research direction for scaling and privacy aspects.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

HydraGNN v4.0

The new version of HydraGNN v4.0 provides additional core capabilities, such as: Inclusion of multi-body atomistic cluster expansion MACE, polarizable atom interaction neural network PAINN, and equivariant principal neighborhood aggregation (PNAEq) among the message passing layers supported -Inclusion of graph transformers to directly model long-range interactions between nodes that are distant in the graph topology Integration of graph transformers with message passing layers by combining the graph embedding generated by the two mechanisms, which allows for an improved expressivity of the HydraGNN architecture Improved re-implementation of multi-task learning (MTL) to allow its use for stabilized training across imbalanced, multi-source, multi-fidelity data Introduction of multi-task parallelism, a newly proposed type of model parallelism specifically for MTL architectures, which allows to dispatch different output decoding heads to different GPU devices Integration of multi-task parallelism with pre-existing distributed data parallelism to enable a 2D parallelization for distributed training Improved portability of the distributed training across Intel GPUs, which has been testes on ALCF exascale supercomputer Aurora Inclusion of 2-level fine-grained energy profilers portable across NVIDIA, AMD, and Intel GPUs to monitor the power and energy consumption associated with different functions executed by the HydraGNN code during data pre-load and training Restructuring of previous examples and inclusion of new sets of examples to illustrate the download, preprocess, and training of HydraGNN models on new large-scale open-source datasets for atomistic materials modeling (e.g., Alexandria, Transition1x, OMat24, OMol25)

Lupo Pasini, Massimiliano [Oak Ridge National Labo↗
DOE OSTI2025

Optimizing temperature distributions for training neural quantum states using parallel tempering

Parametrized artificial neural networks (ANNs) can be very expressive ansatzes for variational algorithms, reaching state-of-the-art energies on many quantum many-body Hamiltonians. Nevertheless, the training of the ANN can be slow and stymied by the presence of local minima in the parameter landscape. One approach to mitigate this issue is to use parallel tempering methods, and in this work, we focus on the role played by the temperature distribution of the parallel tempering replicas. Using an adaptive method that adjusts the temperatures in order to equate the exchange probability between neighboring replicas, we show that this temperature optimization can significantly increase the success rate of the variational algorithm with negligible computational cost by eliminating bottlenecks in the replicas' random walk. Furthermore, we demonstrate this using two different neural networks, a restricted Boltzmann machine and a feedforward network, which we use to study a toy problem based on a permutation invariant Hamiltonian with a pernicious local minimum and the 𝐽 1 −𝐽 2 model on a rectangular lattice.

Neural network simulations↗
DOE OSTI2025

5G integrated edge computing platform for efficient component monitoring in coal-fired power plants

This project developed a cutting-edge 5G-integrated edge computing framework to enhance operational efficiency and reliability in coal-fired power plants through real-time component monitoring and anomaly detection. The initiative focused on leveraging distributed machine learning, federated learning, and 5G-based dynamic network slicing to support scalable, fault-tolerant monitoring environments to meet the operational requirements in industrial control systems. With a Distributed Edge Computing Service (DECS) orchestration, this project enabled federated learning at edge for condition monitoring and introduced adaptive client selection strategies to minimize communication overhead. Scalable distributed training was achieved using the Horovod framework, thus enhancing performance across edge nodes. In the realm of 5G networking, the project designed and deployed reconfigurable, QoS-aware network slicing tailored for operational technology (OT) environments, integrating software-defined networks to bolster cyber-resilience and enabling dynamic slicing for federated learning workloads. A significant milestone was the development of a virtualized ICS environment with 5G core integration—which allowed elastic and fault tolerant distributed training on real-world datasets such as NASA Bearings, Hydraulic Systems, and TEP. To broaden the impact of the project, a TRL-3 virtualized ICS testbed for research and education was designed. This project engaged several graduate and undergraduate students to conduct research on the cutting-edge technology, and it resulted in one PhD dissertation, one MS thesis, and over 14 peer-reviewed publications. With the support of this project students also participated in national cybersecurity competitions to improve their professional development skills.

20 FOSSIL-FUELED POWER PLANTS↗
DOE OSTI2026

Active learning enables generation of molecules that advance the known Pareto front

Although generative models hold promise for discovering molecules with optimized desired properties, they often fail to suggest synthesizable molecules that improve upon the properties of the structures represented in the training distribution. We find that this limitation arises not only from the molecule generation process itself, but also from the poor generalization capabilities of molecular property predictors. We address this challenge by creating a closed-loop molecule generation pipeline with iterative retraining on new quantum chemical simulation data. Compared against static, single-pass generative modeling approaches, only our closed-loop iterative workflow generates molecules with properties extending beyond the training distribution (up to 0.44 standard deviations beyond the original range) and achieves a 79% improvement in out-of-distribution molecule classification accuracy. Furthermore, by conditioning molecular generation on thermodynamic stability data obtained during the iterative loop, the proportion of stable and hence potentially synthesizable molecules generated is 3.5x higher than the next-best model.

Chemistry↗
DOE OSTI2022

Physics-assisted generative adversarial network for X-ray tomography

X-ray tomography is capable of imaging the interior of objects in three dimensions non-invasively, with applications in biomedical imaging, materials science, electronic inspection, and other fields. The reconstruction process can be an ill-conditioned inverse problem, requiring regularization to obtain satisfactory results. Recently, deep learning has been adopted for tomographic reconstruction. Unlike iterative algorithms which require a distribution that is known a priori , deep reconstruction networks can learn a prior distribution through sampling the training distributions. In this work, we develop a Physics-assisted Generative Adversarial Network (PGAN), a two-step algorithm for tomographic reconstruction. In contrast to previous efforts, our PGAN utilizes maximum-likelihood estimates derived from the measurements to regularize the reconstruction with both known physics and the learned prior. Compared with methods with less physics assisting in training, PGAN can reduce the photon requirement with limited projection angles to achieve a given error rate. The advantages of using a physics-assisted learned prior in X-ray tomography may further enable low-photon nanoscale imaging.

47 OTHER INSTRUMENTATION↗
DOE OSTI2022

Distributed deep learning training using silicon photonic switched architectures

The scaling trends of deep learning models and distributed training workloads are challenging network capacities in today’s datacenters and high-performance computing (HPC) systems. We propose a system architecture that leverages silicon photonic (SiP) switch-enabled server regrouping using bandwidth steering to tackle the challenges and accelerate distributed deep learning training. In addition, our proposed system architecture utilizes a highly integrated operating system-based SiP switch control scheme to reduce implementation complexity. To demonstrate the feasibility of our proposal, we built an experimental testbed with a SiP switch-enabled reconfigurable fat tree topology and evaluated the network performance of distributed ring all-reduce and parameter server workloads. The experimental results show up to 3.6× improvements over the static non-reconfigurable fat tree. Our large-scale simulation results show that server regrouping can deliver up to 2.3× flow throughput improvement for a 2× tapered fat tree and a further 11% improvement when higher-layer bandwidth steering is employed. The collective results show the potential of integrating SiP switches into datacenters and HPC systems to accelerate distributed deep learning training.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2022

AI and extreme scale computing to learn and infer the physics of higher order gravitational wave modes of quasi-circular, spinning, non-precessing black hole mergers

We use artificial intelligence (AI) to learn and infer the physics of higher order gravitational wave modes of quasi-circular, spinning, non precessing binary black hole mergers. We trained AI models using 14 million waveforms, produced with the surrogate model NRHybSur3dq8, that include modes up to $\ell$ ≤ 4 and (5,5), except for (4,0) and (4,1), that describe binaries with mass-ratios $\textit{q}$ ≤ 8, individual spins $s^z_{\{1,2\}} \in$[–0.8,0.8], and inclination angle $θ \in$ [0,π]. Our probabilistic AI surrogates can accurately constrain the mass-ratio, individual spins, effective spin, and inclination angle of numerical relativity waveforms that describe such signal manifold. We compared the predictions of our AI models with Gaussian process regression, random forest, k-nearest neighbors, and linear regression, and with traditional Bayesian inference methods through the PyCBC Inference toolkit, finding that AI outperforms all these approaches in terms of accuracy, and are between three to four orders of magnitude faster than traditional Bayesian inference methods. Our AI surrogates were trained within 3.4 hours using distributed training on 1,536 NVIDIA V100 GPUs in the Summit supercomputer.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Evaluation of pre-training large language models on leadership-class supercomputers

Large language models (LLMs) have arisen rapidly to the center stage of artificial intelligence as the foundation models applicable to many downstream learning tasks. However, how to effectively build, train, and serve such models for many high-stake and first-principle-based scientific use cases are both of great interests and of great challenges. Moreover, pre-training LLMs with billions or even trillions of parameters can be prohibitively expensive not just for academic institutions, but also for well-funded industrial and government labs. Furthermore, the energy cost and the environmental impact of developing LLMs must be kept in mind. Here, in this work, we conduct a first-of-its-kind performance analysis to understand the time and energy cost of pre-training LLMs on the Department of Energy (DOE)’s leadership-class supercomputers. Employing state-of-the-art distributed training techniques, we evaluate the computational performance of various parallelization approaches at scale for a range of model sizes, and establish a projection model for the cost of full training. Our findings provide baseline results, best practices, and heuristics for pre-training such large models that should be valuable to HPC community at large. We also offer insights and optimization strategies for using the first exascale computing system, Frontier, to train models of the size of GPT-3 and beyond.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

A Scalable Pipeline for Gigapixel Whole Slide Imaging Analysis on Leadership Class HPC Systems

Whole Slide Imaging (WSI) captures microscopic details of a patient's histopathological features at multiple res-olutions organized across different levels. Images produced by WSI are gigapixel-sized, and saving a single image in memory requires a few gigabytes which is scarce since a complicated model occupies tens of gigabytes. Performing a simple met-ric operation on these large images is also expensive. High-performance computing (HPC) can help us quickly analyze such large images using distributed training of complex deep learning models. One popular approach in analyzing these images is to divide a WSI image into smaller tiles (patches) and then train a simpler model with these reduced-sized but large numbers of patches. However, we need to solve three pre-processing challenges efficiently for pursuing this patch-based approach. 1) Creating small patches from a high-resolution image can result in a high number (hundreds of thousands per image) of patches. Storing and processing these images can be challenging due to a large number of I/O and arithmetic operations. To reduce I/Oand memory accesses, an optimal balance between the size and number of patches must exist to reduce I/O and memory accesses. 2) WSI images may have tiny annotated regions for cancer tissue and a significant portion with normal and fatty tissues; correct patch sampling should avoid dataset imbalance. 3) storing and retrieving many patches to and from disk storage might incur I/O latency while training a deep learning model. An efficient distributed data loader should reduce I/O latency during the training and inference steps. This paper explores these three challenges and provides empirical and algorithmic solutions deployed on the Summit supercomputer hosted at the Oak Ridge Leadership Computing Facility.

Dash, Sajal↗