Fast Imprecise Sets: Performance-Driven Floating-Point Mixed-Precision Tuning for GPU Programs
Explore the source record for details and available documents.
SEARCH · Engineering Papers
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.
Explore the source record for details and available documents.
We apply auto-tuning to a hybrid MPI-pthreads lattice Boltzmann computation running on the Cray XT4 at National Energy Research Scientific Computing Center (NERSC). Previous work showed that multicore-specific auto-tuning can improve the performance of lattice Boltzmann magnetohydrodynamics (LBMHD) by a factor of 4x when running on dual- and quad-core Opteron dual-socket SMPs. We extend these studies to the distributed memory arena via a hybrid MPI/pthreads implementation. In addition to conventional auto-tuning at the local SMP node, we tune at the message-passing level to determine the optimal aspect ratio as well as the correct balance between MPI tasks and threads per MPI task. Our study presents a detailed performance analysis when moving along an isocurve of constant hardware usage: fixed total memory, total cores, and total nodes. Overall, our work points to approaches for improving intra- and inter-node efficiency on large-scale multicore systems for demanding scientific applications.
Most navigation systems currently operated by NASA are ground-based, and require extensive support to produce accurate results. Recently developed systems that use Kalman filter and GPS (Global Positioning Systems) data for orbit determination greatly reduce dependency on ground support, and have potential to provide significant economies for NASA spacecraft navigation. These systems, however, still rely on manual tuning from analysts. A sophisticated neuro-fuzzy component fully integrated with the flight navigation system can perform the self-tuning capability for the Kalman filter and help the navigation system recover from estimation errors in real time.
Most navigation systems currently operated by NASA are ground-based, and require extensive support to produce accurate results. Recently developed systems that use Kalman filter and GPS data for orbit determination greatly reduce dependency on ground support, and have potential to provide significant economies for NASA spacecraft navigation. These systems, however, still rely on manual tuning from analysts. A sophisticated neuro-fuzzy component fully integrated with the flight navigation system can perform the self-tuning capability for the Kalman filter and help the navigation system recover from estimation errors in real time.
As scientific simulations scale to use petascale machines and beyond, the data volumes generated pose a dual problem. First, with increasing machine sizes, the careful tuning of IO routines becomes more and more important to keep the time spent in IO acceptable. It is not uncommon, for instance, to have 20% of an application's runtime spent performing IO in a 'tuned' system. Careful management of the IO routines can move that to 5% or even less in some cases. Second, the data volumes are so large, on the order of 10s to 100s of TB, that trying to discover the scientifically valid contributions requires assistance at runtime to both organize and annotate the data. Waiting for offline processing is not feasible due both to the impact on the IO system and the time required. To reduce this load and improve the ability of scientists to use the large amounts of data being produced, new techniques for data management are required. First, there is a need for techniques for efficient movement of data from the compute space to storage. These techniques should understand the underlying system infrastructure and adapt to changing system conditions. Technologies include aggregation networks, data staging nodes for a closer parity to the IO subsystem, and autonomic IO routines that can detect system bottlenecks and choose different approaches, such as splitting the output into multiple targets, staggering output processes. Such methods must be end-to-end, meaning that even with properly managed asynchronous techniques, it is still essential to properly manage the later synchronous interaction with the storage system to maintain acceptable performance. Second, for the data being generated, annotations and other metadata must be incorporated to help the scientist understand output data for the simulation run as a whole, to select data and data features without concern for what files or other storage technologies were employed. All of these features should be attained while maintaining a simple deployment for the science code and eliminating the need for allocation of additional computational resources.
In a hydrogen station that operates with direct fueling through the use of a 700 bar boost compressor, the outlet hydrogen temperature can significantly increase, stressing the chiller system. Here, this paper evaluates improvements that can be made to the auxiliary cooling system integrated with the compressor. Both theoretical modeling and experiments were performed at Cal State LA Hydrogen Research and Fueling Facility. The findings suggest that adjusting the auxiliary closed-loop cooling system from 15 °C to 10 °C reduced the station energy consumption and decreased the demand on the station chiller that needed to provide –20 °C hydrogen at the hose. The overall energy consumption for a single fueling reduced by between 2.86 and 9.43% for the set of experiments conducted. After the temperature of the closed-loop cooling system was reduced by 5 °C, the boost compressor outlet temperature dropped from 46-50 °C–40 °C and consequently at the hose the hydrogen temperature declined by 3 °C. Results were scaled up with a forecast on the number of daily refueling events. With a low number of daily fuelings, the proposed set-up showed a minor influence on the overall station energy consumption. However, the benefits were more pronounced for a connector station with sales at 180 kg/day, where the energy efficiency improved by between 1.4 and 5.5%, and even more so for a higher capacity station at 360 kg/day, where the improvement was between 2.9 and 8%.
The current generation of large language models (LLMs) has limited chemical knowledge. Recently, it has been shown that these LLMs can learn and predict chemical properties through fine-tuning. Using natural language to train machine learning models opens doors to a wider chemical audience, as field-specific featurization techniques can be omitted. In this work, we explore the potential and limitations of this approach. We studied the performance of fine-tuning three open-source LLMs (GPT-J-6B, Llama-3.1-8B, and Mistral-7B) for a range of different chemical questions. We benchmark their performances against “traditional” machine learning models and find that, in most cases, the fine-tuning approach is superior for a simple classification problem. Depending on the size of the dataset and the type of questions, we also successfully address more sophisticated problems. The most important conclusions of this work are that, for all datasets considered, their conversion into an LLM fine-tuning training set is straightforward and that fine-tuning with even relatively small datasets leads to predictive models. These results suggest that the systematic use of LLMs to guide experiments and simulations will be a powerful technique in any research study, significantly reducing unnecessary experiments or computations.
This article introduces a new postsynthetic modification method for tuning the separation properties of carbon molecular sieve (CMS) membranes. Polymeric hollow fibers of 6FDA/BPDA-DAM were pyrolyzed to 550°C under inert Argon, and were then exposed to a tetrahydrofuran solution containing PPM levels of paraphenylenediamine dopant. The original goal of the treatment was to modify the morphology in a manner to prevent relaxation of the CMS to suppress physical aging, envisioned to be analogous to that in glassy polymer membranes where reduction of “unrelaxed free volume” causes reduced permeance with a mild increase in selectivity. Results of long-term CO 2 and CH 4 permeation experiments of doped fibers showed that aging-resistant membranes did not result. This fact notwithstanding, results revealed this approach provides a valuable new tuning tool for CMS transport properties. Complementary gas sorption experiments performed suggest the amine dopants tune large ultramicropores, thereby increasing size and shape based diffusion selectivity.
Traditional analysis of highly distorted micro X‐ray diffraction (μ‐XRD) patterns from hydrothermal fluid environments is a time‐consuming process, often requiring substantial data preprocessing and labeled experimental data. Herein, the potential of deep learning with a multitask learning (MTL) architecture to overcome these limitations is demonstrated. MTL models are trained to identify phase information in μ‐XRD patterns, minimizing the need for labeled experimental data and masking preprocessing steps. Notably, MTL models show superior accuracy compared to binary classification convolutional neural networks. Additionally, introducing a tailored cross‐entropy loss function improves MTL model performance. Most significantly, MTL models tuned to analyze raw and unmasked XRD patterns achieve close performance to models analyzing preprocessed data, with minimal accuracy differences. This work indicates that advanced deep learning architectures like MTL can automate arduous data handling tasks, streamline the analysis of distorted XRD patterns, and reduce the reliance on labor‐intensive experimental datasets.
Redox-active porous polymers are becoming promising organic electrode materials for alkali-ion batteries because of their tunable pore size, flexible chemical structure, and high thermal stability. This work presents various porous bipolar polymers as cathodes in Na-ion batteries (NIBs) and K-ion batteries (KIBs). Two structural units containing carbonyl and amine active centers, respectively, were introduced into the repeating units of the polymers by the polycondensation of N,N,N',N'-tetrakis(4-aminophenyl)-1,4-phenylenediamine (TAP) and various dianhydrides. Here, the resulting bipolar polymers undergo multi-electron redox reactions involving both cations and anions during battery charge and discharge. The impacts of extended conjugation structures and porosity on the electrochemical performances were investigated by tuning the conjugation structures in the dianhydride monomers. The bipolar polymers with extended conjugation structures and small pore size exhibit superior electrochemical performance in NIBs and KIBs in terms of high redox potentials, long cycle life, and fast-charging capability, demonstrating great promise as organic cathode materials for alkali-ion batteries.
An emerging approach in the field of aircraft engine controls and system health management is the inclusion of real-time, onboard models for the inflight estimation of engine performance variations. This technology, typically based on Kalman-filter concepts, enables the estimation of unmeasured engine performance parameters that can be directly utilized by controls, prognostics, and health-management applications. A challenge that complicates this practice is the fact that an aircraft engine s performance is affected by its level of degradation, generally described in terms of unmeasurable health parameters such as efficiencies and flow capacities related to each major engine module. Through Kalman-filter-based estimation techniques, the level of engine performance degradation can be estimated, given that there are at least as many sensors as health parameters to be estimated. However, in an aircraft engine, the number of sensors available is typically less than the number of health parameters, presenting an under-determined estimation problem. A common approach to address this shortcoming is to estimate a subset of the health parameters, referred to as model tuning parameters. The problem/objective is to optimally select the model tuning parameters to minimize Kalman-filterbased estimation error. A tuner selection technique has been developed that specifically addresses the under-determined estimation problem, where there are more unknown parameters than available sensor measurements. A systematic approach is applied to produce a model tuning parameter vector of appropriate dimension to enable estimation by a Kalman filter, while minimizing the estimation error in the parameters of interest. Tuning parameter selection is performed using a multi-variable iterative search routine that seeks to minimize the theoretical mean-squared estimation error of the Kalman filter. This approach can significantly reduce the error in onboard aircraft engine parameter estimation applications such as model-based diagnostic, controls, and life usage calculations. The advantage of the innovation is the significant reduction in estimation errors that it can provide relative to the conventional approach of selecting a subset of health parameters to serve as the model tuning parameter vector. Because this technique needs only to be performed during the system design process, it places no additional computation burden on the onboard Kalman filter implementation. The technique has been developed for aircraft engine onboard estimation applications, as this application typically presents an under-determined estimation problem. However, this generic technique could be applied to other industries using gas turbine engine technology.
As scientific and engineering domains attempt to effectively analyze the deluge of data arriving from sensors and instruments, machine learning is becoming a key data mining tool to build prediction models. Regression tree is a popular learning model that combines decision trees and linear regression to forecast numerical target variables based on a set of input features. Map Reduce is well suited for addressing such data intensive learning applications, and a proprietary regression tree algorithm, PLANET, using MapReduce has been proposed earlier. In this paper, we describe an open source implement of this algorithm, OpenPlanet, on the Hadoop framework using a hybrid approach. Further, we evaluate the performance of OpenPlanet using realworld datasets from the Smart Power Grid domain to perform energy use forecasting, and propose tuning strategies of Hadoop parameters to improve the performance of the default configuration by 75% for a training dataset of 17 million tuples on a 64-core Hadoop cluster on FutureGrid.
To meet PIP-II upgrade requirements, Fermilab Booster losses need to be reduced by 50% compared to present levels. So far, simulations are not good enough to predict loss patterns. Thus, an extensive Booster tune up will be necessary to achieve required performance. In this paper we present an effort to build a data-driven loss model using Bayesian techniques, and subsequently to rebalance losses for higher trip margins. We first created several sets of spatially and temporally isolated orbit and optics knobs, and trained Gaussian process models for each beam loss monitor as well as beam current. Novel techniques of uncertainty constraints and approximate GP fitting were introduced to handle safety and timing requirements. We then performed single and multi-objective tuning using scalarized objectives comprised of critical beam loss locations. We achieved significant rebalancing of losses, increasing margins by 25%, as well as an overall improvement in transmission efficiency of 0.4%. Automated data collection is being developed so that more accurate surrogate models can be trained over time.
High performance control design for a flexible space structure is challenging since high fidelity plant models are difficult to obtain a priori. Uncertainty in the control design models typically require a very robust, low performance control design which must be tuned on-orbit to achieve the required performance. A new procedure for refining a multivariable open loop plant model based on closed-loop response data is presented. Using a minimal representation of the state space dynamics, a least squares prediction error method is employed to estimate the plant parameters. This control-relevant system identification procedure stresses the joint nature of the system identification and control design problem by seeking to obtain a model that minimizes the difference between the predicted and actual closed-loop performance. This paper presents an algorithm for iterative closed-loop system identification and controller redesign along with illustrative examples.
High performance control design for a flexible space structure is challenging since high fidelity plant models are di.cult to obtain a priori. Uncertainty in the control design models typically require a very robust, low performance control design which must be tuned on-orbit to achieve the required performance. Closed loop system identi.cation is often required to obtain a multivariable open loop plant model based on closed-loop response data. In order to provide an accurate initial plant model to guarantee convergence for standard local optimization methods, this paper presents a global parameter optimization method using genetic algorithms. A minimal representation of the state space dynamics is employed to mitigate the non-uniqueness and over-parameterization of general state space realizations. This control-relevant system identi.cation procedure stresses the joint nature of the system identi.cation and control design problem by seeking to obtain a model that minimizes the di.erence between the predicted and actual closed-loop performance.
Particle accelerators are vital tools in physics, medicine, and industry, requiring precise tuning to ensure optimal beam performance. However, real-world deviations from idealized simulations make beam tuning a time-consuming and error-prone process. In this work, we propose an explanation-driven framework for providing actionable insight into beamline operations, with a focus on the injector beamline at the Continuous Electron Beam Accelerator Facility (CEBAF). We represent beamline configurations as heterogeneous graphs, where setting nodes represent elements that human operators can actively adjust during beam tuning, and reading nodes passively provide diagnostic feedback. To identify the most influential setting nodes responsible for differences between any two beamline configurations, our approach first predicts the resulting changes in reading nodes caused by variations in settings, and then learns importance scores that capture the joint influence of multiple setting nodes. Experimental results on real-world CEBAF injector data demonstrate the framework’s ability to generate interpretable insights that can assist human operators in beamline tuning and reduce operational overhead.
This work introduces a generalized framework for automatically tuning stencil computations to achieve superior performance on a broad range of multicore architectures. Stencil (nearest-neighbor) based kernels constitute the core of many important scientific applications involving block-structured grids. Auto-tuning systems search over optimization strategies to find the combination of tunable parameters that maximizes computational efficiency for a given algorithmic kernel. Although the auto-tuning strategy has been successfully applied to libraries, generalized stencil kernels are not amenable to packaging as libraries. Studied kernels in this work include both memory-bound kernels as well as a computation-bound bilateral filtering kernel. We introduce a generalized stencil auto-tuning framework that takes a straightforward Fortran expression of a stencil kernel and automatically generates tuned implementations of the kernel in C or Fortran to achieve performance portability across diverse computer architectures.
One way of imaging X-ray emission from solar flares is to measure Fourier components of the spatial X-ray source distribution. We present a new compressed sensing-based algorithm named VIS-CS, which reconstructs the spatial distribution from such Fourier components. We demonstrate the application of the algorithm on synthetic and observed solar flare X-ray data from the Reuven Ramaty High Energy Solar Spectroscopic Imager satellite and compare its performance with existing algorithms. VIS-CS produces competitive results with accurate photometry and morphology, without requiring any algorithm- and X-ray-source-specific parameter tuning. Its robustness and performance make this algorithm ideally suited for the generation of quicklook images or large image cubes without user intervention, such as for imaging spectroscopy analysis.