Search NASA⌕ Search

SEARCH · Search NASA

Results for “memory efficiency”

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 1,567 records · Page 87

Evaluating the Potential and Challenges of an Uncertainty Quantification Method for Long Short–Term Memory Models for Soil Moisture Predictions

Recently, recurrent deep networks have shown promise to harness newly available satellite–sensed data for long–term soil moisture projections. However, to be useful in forecasting, deep networks must also provide uncertainty estimates. Here we evaluated Monte Carlo dropout with an input–dependent data noise term (MCD+N), an efficient uncertainty estimation framework originally developed in computer vision, for hydrologic time series predictions. MCD+N simultaneously estimates a heteroscedastic input–dependent data noise term (a trained error model attributable to observational noise) and a network weight uncertainty term (attributable to insufficiently constrained model parameters). Although MCD+N has appealing features, many heuristic approximations were employed during its derivation, and rigorous evaluations and evidence of its asserted capability to detect dissimilarity were lacking. To address this, we provided an in–depth evaluation of the scheme's potential and limitations. We showed that for reproducing soil moisture dynamics recorded by the Soil Moisture Active Passive (SMAP) mission, MCD+N indeed gave a good estimate of predictive error, provided that we tuned a hyperparameter and used a representative training data set. The input–dependent term responded strongly to observational noise, while the model term clearly acted as a detector for physiographic dissimilarity from the training data, behaving as intended. However, when the training and test data were characteristically different, the input–dependent term could be misled, undermining its reliability. Additionally, due to the data–driven nature of the model, data noise also influences network weight uncertainty, and therefore the two uncertainty terms are correlated. Altogether, this approach has promise, but care is needed to interpret the results.

54 ENVIRONMENTAL SCIENCES↗

Long short-term memory embedded nudging schemes for nonlinear data assimilation of geophysical flows

Reduced rank nonlinear filters are increasingly utilized in data assimilation of geophysical flows, but often require a set of ensemble forward simulations to estimate forecast covariance. On the other hand, predictor-corrector type nudging approaches are still attractive due to their simplicity of implementation when more complex methods need to be avoided. However, optimal estimate of nudging gain matrix might be cumbersome. In this paper, we put forth a fully nonintrusive recurrent neural network approach based on a long short-term memory (LSTM) embedding architecture to estimate the nudging term, which plays a role not only to force the state trajectories to the observations but also acts as a stabilizer. Furthermore, our approach relies on the power of archival data and the trained model can be retrained effectively due to power of transfer learning in any neural network applications. In order to verify the feasibility of the proposed approach, we perform twin experiments using Lorenz 96 system. Our results demonstrate that the proposed LSTM nudging approach yields more accurate estimates than both extended Kalman filter (EKF) and ensemble Kalman filter (EnKF) when only sparse observations are available. With the availability of emerging AI-friendly and modular hardware technologies and heterogeneous computing platforms, we articulate that our simplistic nudging framework turns out to be computationally more efficient than either the EKF or EnKF approaches.

42 ENGINEERING↗

DISP: Optimizations towards Scalable MPI Startup

Despite the popularity of MPI for high performance computing, the startup of MPI programs faces a scalability challenge as both the execution time and memory consumption increase drastically at scale. We have examined this problem using the collective modules of Cheetah and Tuned in Open MPI as representative implementations. Previous improvements for collectives have focused on algorithmic advances and hardware off-load. In this paper, we examine the startup cost of the collective module within a communicator and explore various techniques to improve its efficiency and scalability. Accordingly, we have developed a new scalable startup scheme with three internal techniques, namely Delayed Initialization, Module Sharing and Prediction-based Topology Setup (DISP). Our DISP scheme greatly benefits the collective initialization of the Cheetah module. At the same time, it helps boost the performance of non-collective initialization in the Tuned module. We evaluate the performance of our implementation on Titan supercomputer at ORNL with up to 4096 processes. The results show that our delayed initialization can speed up the startup of Tuned and Cheetah by an average of 32.0% and 29.2%, respectively, our module sharing can reduce the memory consumption of Tuned and Cheetah by up to 24.1% and 83.5%, respectively, and our prediction-based topology setup can speed up the startup of Cheetah by up to 80%.

Fu, Huansong↗

A Step towards Energy Efficient Computing: Redesigning a Hydrodynamic Application on CPU-GPU, In: 2014 IEEE 28th International Parallel and Distributed Processing Symposium

Power and energy consumption are becoming an increasing concern in high performance computing. Compared to multi-core CPUs, GPUs have a much better performance per watt. In this paper we discuss efforts to redesign the most computation intensive parts of BLAST, an application that solves the equations for compressible hydrodynamics with high order finite elements, using GPUs BLAST, Dobrev. In order to exploit the hardware parallelism of GPUs and achieve high performance, we implemented custom linear algebra kernels. We intensively optimized our CUDA kernels by exploiting the memory hierarchy, which exceed the vendor's library routines substantially in performance. We proposed an auto tuning technique to adapt our CUDA kernels to the orders of the finite element method. Compared to a previous base implementation, our redesign and optimization lowered the energy consumption of the GPU in two aspects: 60% less time to solution and 10% less power required. Compared to the CPU-only solution, our GPU accelerated BLAST obtained a 2.5× overall speedup and 1.42× energy efficiency (green up) using 4th order (Q_4) finite elements, and a 1.9× speedup and 1.27× green up using 2nd order (Q2) finite elements.

Computer Science↗

Refining HPCToolkit for application performance analysis at exascale

As part of the US Department of Energy’s Exascale Computing Project (ECP), Rice University has been refining its HPCToolkit performance tools to better support measurement and analysis of applications executing on exascale supercomputers. To efficiently collect performance measurements of GPU-accelerated applications, HPCToolkit employs novel non-blocking data structures to communicate performance measurements between tool threads and application threads. To attribute performance information in detail to source lines, loop nests, and inlined call chains, HPCToolkit performs parallel analysis of large CPU and GPU binaries involved in the execution of an exascale application to rapidly recover mappings between machine instructions and source code. To analyze terabytes of performance measurements gathered during executions at exascale, HPCToolkit employs distributed-memory parallelism, multithreading, sparse data structures, and out-of-core streaming analysis algorithms. To support interactive exploration of profiles up to terabytes in size, HPCToolkit’s hpcviewer graphical user interface uses out-of-core methods to visualize performance data. The result of these efforts is that HPCToolkit now supports collection, analysis, and presentation of profiles and traces of GPU-accelerated applications at exascale. These improvements have enabled HPCToolkit to efficiently measure, analyze and explore terabytes of performance data for executions using as many as 64K MPI ranks and 64K GPU tiles on ORNL’s Frontier supercomputer. HPCToolkit’s support for measurement and analysis of GPU-accelerated applications has been employed to study a collection of open-science applications developed as part of ECP. This paper reports on these experiences, which provided insight into opportunities for tuning applications, strengths and weaknesses of HPCToolkit itself, as well as unexpected behaviors in executions at exascale.

Adhianto, Laksono↗

Automatic Code Generation for High-Performance Graph Algorithms

Graph problems are common across fields of scientific computing and social sciences. However, despite their importance, implementing graph algorithms effectively on modern computing systems is a challenging task that requires significant programming effort and generally results in customized implementations. Current computing and memory hierarchies are not architected for irregular computations resulting in challenges for graph algorithms to achieve high performance on those architectures. In this paper, we present GraphX, a novel compiler framework and DSL designed to simplify the development of efficient graph algorithms and achieve high performance on modern computing systems. GraphX consists of a DSL for efficient implementation of graph algorithms, various optimizations, such as support for sparse linear algebra and workspace transformations, optimized graph primitives, including semiring and masking, and a high-performance code generation engine. Using GraphX, users can implement graph algorithms using a semantically-rich language with graph-oriented operators. GraphX uses these semantics to automatically generate efficient code for target architectures, increasing performance and portability across architectures. The composable nature of GraphX makes it possible to extend the set of optimizations and architectures without modifying the source code. We demonstrate GraphX outperforms state-of-the-art graph libraries, such as LAGraph, up to $3.7 speedup in semiring operations, $2.19 speedup in an important sparse computational kernel, and $9.05 speedup in graph processing algorithms.

compiler, graph algorithms, semiring, masking, wor↗

Sequence length scaling in vision transformers for scientific images on frontier

Vision Transformers (ViTs) are pivotal for foundational models in scientific imagery, including Earth science applications, due to their capability to process large sequence lengths. While transformers for text have inspired scaling sequence lengths in ViTs, adapting these for ViTs introduces unique challenges. We develop distributed sequence parallelism for ViTs, enabling them to handle up to 1M tokens. Our approach, leveraging DeepSpeed-Ulysses and Long-Sequence-Segmentation with model sharding, is the first to apply sequence parallelism in ViT training, achieving a 94% batch scaling efficiency on 2,048 AMD-MI250X GPUs. Evaluating sequence parallelism in ViTs, particularly in models up to 10B parameters, highlighted substantial bottlenecks. We countered these with hybrid sequence, pipeline, and flash attention strategies, to scale beyond single GPU memory limits. Our method significantly enhances climate modeling accuracy by 20% in temperature predictions, marking the first training of a vision transformer model to convergence with a sequence length of 188K tokens, using full self-attention.

Tsaris, Aristeidis (aris) [ORNL] (ORCID:0000000277↗

MOOSE Framework Speed and Memory Optimizations

The Multiphysics Object Oriented Simulation Environment (MOOSE) framework has been developed to solve a variety of complex simulations ranging from groundwater flows to full reactor simulations. MOOSE excels at solving fully-implicit, fully-coupled, highly nonlinear models using the Jacobian Free Newton-Krylov (JFNK) method relying on years of expertise in several dependent packages including PETSc, libMesh, Hypre and many others. The modular design of MOOSE enables the rapid development of new physics models that can be coupled with relative ease within the framework for building up simulations of different levels of complexity. With the ongoing adoption of MOOSE to new application areas along with streamlining of existing simulations, the priorities must be adjusted to include code optimizations. This includes both execution speed and the memory footprint of MOOSE-based simulations. These two categories can be further broken down into ''micro-optimizations'' or source code optimizations and algorithmic optimizations where source code is restructured to accomplish the same tasks through a more efficient means. However, before any code optimizations can occur, one must have robust way of measuring these critical simulation properties to identify the best candidates for optimizations along with having a way to identify the effectiveness of proposed optimizations to the framework. This report discusses the preliminary work towards the implementation of a robust execution baseline of the framework and derivative applications under a variety of different simulations. In addition, several potential optimization target areas in both the MOOSE framework and dependent libraries have been identified. An implementation strategy and potential benefits are discussed.

97 MATHEMATICS AND COMPUTING↗

$\mathrm{SageNet}$: Fast Neural Network Emulation of the Stiff-amplified Gravitational Waves from Inflation

Accurate modeling of the inflationary gravitational waves (GWs) requires time-consuming, iterative numerical integrations of differential equations to take into account their backreaction on the expansion history. To improve computational efficiency while preserving accuracy, we present the Stiff-amplified Gravitational-wave Emulator Network (SageNet), a deep learning framework designed to replace conventional numerical solvers (code available at https://github.com/YifangLuo/SageNet). SageNet employs a long short-term memory architecture to emulate the present-day energy density spectrum of the inflationary GWs with possible stiff amplification, Ω GW (f). Trained on a data set of 25,689 numerically generated solutions, SageNet allows accurate reconstructions of Ω GW (f) and generalizes well to a wide range of cosmological parameters; 90.9% of the test emulations with randomly distributed parameters exhibit errors of under 4%. In addition, SageNet demonstrates its ability to learn and reproduce the artificial, adaptive sampling patterns in numerical calculations, which implement denser sampling of frequencies around changes in spectral indices in Ω GW (f). The dual capability of learning both physical and artificial features of the numerical GW spectra establishes SageNet as a robust alternative to exact numerical methods. Finally, our benchmark tests show that SageNet reduces the computation time from tens of seconds to milliseconds, achieving a speedup of ∼10 4 times over standard CPU-based numerical solvers with the potential for further acceleration on GPU hardware. These capabilities make SageNet a powerful tool for accelerating Bayesian inference procedures for extended cosmological models. In a broad sense, the SageNet framework offers a fast, accurate, and generalizable solution to modeling cosmological observables whose theoretical predictions demand costly differential equation solvers.

Astronomy data modeling↗

Suggested architecture for a specialized fluid dynamics computer

Future flow simulations in 3-D will require computers with extremely large main memories and an advantageous ratio between computer cost and arithmetic speed. Since random access memories are very expensive, a pipeline design is proposed which allows the use of much cheaper sequential devices without any sacrifice in speed for vector references (even with arbitrary spacing between successive elements). Also scalar arithmetic can be performed efficiently. The comparatively low speed of the proposed machine (about 10 to the 7th power operations per second) would be offset by a very low price per unit, making mass production possible.

Fornberg, B.↗

Multiprocessing on supercomputers for computational aerodynamics

Little use is made of multiple processors available on current supercomputers (computers with a theoretical peak performance capability equal to 100 MFLOPS or more) to improve turnaround time in computational aerodynamics. The productivity of a computer user is directly related to this turnaround time. In a time-sharing environment, such improvement in this speed is achieved when multiple processors are used efficiently to execute an algorithm. The concept of multiple instructions and multiple data (MIMD) is applied through multitasking via a strategy that requires relatively minor modifications to an existing code for a single processor. This approach maps the available memory to multiple processors, exploiting the C-Fortran-Unix interface. The existing code is mapped without the need for developing a new algorithm. The procedure for building a code utilizing this approach is automated with the Unix stream editor.

Yarrow, Maurice↗

Simple, Lightweight, Thermomechanical Coupling

Self-actuating thermal coupling device is light, and transfers heat efficiently. Coupling is thermostatic, making contact as necessary to maintain nearly steady temperature at hot side of interface. Hot pins on hot plate poised near cold pins on cold-plate. As temperature of hot plate rises, shape-memory wire contracts, drawing hot pin against cold pin. When temperature falls below critical level, wire expands, allowing hot pin to pull away. Critical temperature determined by composition of alloy.

Milam, M. Bruce↗

Molecular implementation of molecular shift register memories

An electronic shift register memory (20) at the molecular level is described. The memory elements are based on a chain of electron transfer molecules (22) and the information is shifted by photoinduced (26) electron transfer reactions. Thus, multi-step sequences of charge transfer reactions are used to move charge with high efficiency down a molecular chain. The device integrates compositions of the invention onto a VLSI substrate (36), providing an example of a molecular electronic device which may be fabricated. Three energy level schemes, molecular implementation of these schemes, optical excitation strategies, charge amplification strategies, and error correction strategies are described.

Beratan, David N.↗

Rao-Blackwellization for Adaptive Gaussian Sum Nonlinear Model Propagation

When dealing with imperfect data and general models of dynamic systems, the best estimate is always sought in the presence of uncertainty or unknown parameters. In many cases, as the first attempt, the Extended Kalman filter (EKF) provides sufficient solutions to handling issues arising from nonlinear and non-Gaussian estimation problems. But these issues may lead unacceptable performance and even divergence. In order to accurately capture the nonlinearities of most real-world dynamic systems, advanced filtering methods have been created to reduce filter divergence while enhancing performance. Approaches, such as Gaussian sum filtering, grid based Bayesian methods and particle filters are well-known examples of advanced methods used to represent and recursively reproduce an approximation to the state probability density function (pdf). Some of these filtering methods were conceptually developed years before their widespread uses were realized. Advanced nonlinear filtering methods currently benefit from the computing advancements in computational speeds, memory, and parallel processing. Grid based methods, multiple-model approaches and Gaussian sum filtering are numerical solutions that take advantage of different state coordinates or multiple-model methods that reduced the amount of approximations used. Choosing an efficient grid is very difficult for multi-dimensional state spaces, and oftentimes expensive computations must be done at each point. For the original Gaussian sum filter, a weighted sum of Gaussian density functions approximates the pdf but suffers at the update step for the individual component weight selections. In order to improve upon the original Gaussian sum filter, Ref. [2] introduces a weight update approach at the filter propagation stage instead of the measurement update stage. This weight update is performed by minimizing the integral square difference between the true forecast pdf and its Gaussian sum approximation. By adaptively updating each component weight during the nonlinear propagation stage an approximation of the true pdf can be successfully reconstructed. Particle filtering (PF) methods have gained popularity recently for solving nonlinear estimation problems due to their straightforward approach and the processing capabilities mentioned above. The basic concept behind PF is to represent any pdf as a set of random samples. As the number of samples increases, they will theoretically converge to the exact, equivalent representation of the desired pdf. When the estimated qth moment is needed, the samples are used for its construction allowing further analysis of the pdf characteristics. However, filter performance deteriorates as the dimension of the state vector increases. To overcome this problem Ref. [5] applies a marginalization technique for PF methods, decreasing complexity of the system to one linear and another nonlinear state estimation problem. The marginalization theory was originally developed by Rao and Blackwell independently. According to Ref. [6] it improves any given estimator under every convex loss function. The improvement comes from calculating a conditional expected value, often involving integrating out a supportive statistic. In other words, Rao-Blackwellization allows for smaller but separate computations to be carried out while reaching the main objective of the estimator. In the case of improving an estimator's variance, any supporting statistic can be removed and its variance determined. Next, any other information that dependents on the supporting statistic is found along with its respective variance. A new approach is developed here by utilizing the strengths of the adaptive Gaussian sum propagation in Ref. [2] and a marginalization approach used for PF methods found in Ref. [7]. In the following sections a modified filtering approach is presented based on a special state-space model within nonlinear systems to reduce the dimensionality of the optimization problem in Ref. [2]. First, the adaptive Gaussian sum propagation is explained and then the new marginalized adaptive Gaussian sum propagation is derived. Finally, an example simulation is presented.

state estimation↗

Systems and methods for tensor scheduling

A technique for efficient scheduling of operations in a program for parallelized execution thereof using a multi-processor runtime environment having two or more processors includes constraining the type or number of loop optimization transforms that may be explored such that memory and processing capacity available for the scheduling task are not exceeded, while facilitating a tradeoff between memory locality, parallelization, and/or data communication between memory modules of the multi-processor runtime environment.

Meister, Benoit J.↗

Evaluation of vectorized Monte Carlo algorithms on GPUs for a neutron Eigenvalue problem

Conventional Monte Carlo (MC) methods for radiation transport computations are 'history-based', which means that one particle history at a time is tracked. Simulations based on such methods suffer from thread divergence on the graphics processing unit (GPU), which severely affects the performance of GPUs. To circumvent this limitation, event-based vectorized MC algorithms can be utilized. A versatile software test-bed, called ARCHER - Accelerated Radiation-transport Computations in Heterogeneous Environments - was used for this study. ARCHER facilitates the development and testing of a MC code based on the vectorized MC algorithm implemented on GPUs by using NVIDIA's Compute Unified Device Architecture (CUDA). The ARCHER{sub GPU} code was designed to solve a neutron eigenvalue problem and was tested on a NVIDIA Tesla M2090 Fermi card. We found that although the vectorized MC method significantly reduces the occurrence of divergent branching and enhances the warp execution efficiency, the overall simulation speed is ten times slower than the conventional history-based MC method on GPUs. By analyzing detailed GPU profiling information from ARCHER, we discovered that the main reason was the large amount of global memory transactions, causing severe memory access latency. Several possible solutions to alleviate the memory latency issue are discussed. (authors)

97 MATHEMATICS AND COMPUTING↗

Multiscale high-order/low-order (HOLO) algorithms and applications

We review the state of the art in the formulation, implementation, and performance of so-called high-order/low-order (HOLO) algorithms for challenging multiscale problems. HOLO algorithms attempt to couple one or several high-complexity physical models (the high-order model, HO) with low-complexity ones (the low-order model, LO). The primary goal of HOLO algorithms is to achieve nonlinear convergence between HO and LO components while minimizing memory footprint and managing the computational complexity in a practical manner. Key to the HOLO approach is the use of the LO representations to address temporal stiffness, effectively accelerating the convergence of the HO/LO coupled system. The HOLO approach is broadly underpinned by the concept of nonlinear elimination, which enables segregation of the HO and LO components in ways that can effectively use heterogeneous architectures. The accuracy and efficiency benefits of HOLO algorithms are demonstrated with specific applications to radiation transport, gas dynamics, plasmas (both Eulerian and Lagrangian formulations), and ocean modeling. Across this broad application spectrum, HOLO algorithms achieve significant accuracy improvements at a fraction of the cost compared to conventional approaches. It follows that HOLO algorithms hold significant potential for high-fidelity system scale multiscale simulations leveraging exascale computing.

97 MATHEMATICS AND COMPUTING↗

Efficient single-level solution of hierarchical problems in structural optimization

A demonstration is made of the use of single-level hierarchical solution techniques for engineering design, which, unlike the formulation stage-applied multilevel techniques, are applied at the solution stage. A single-level decomposition technique is developed which reduces computational costs and memory requirements without incurring the disadvantages of multilevel optimization. Savings of 75 percent are obtained in a case involving 500 design variables and 2408 constraints.

Thareja, Rajiv R.↗