Search NASA⌕ Search

SEARCH · Search NASA

Results for “cloud computing, data cloud”

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 739 records · Page 41
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Modern deep neural networks for Direct Normal Irradiance forecasting: A classification approach

The escalating energy demand and the adverse environmental impacts of fossil-fuel use necessitate a shift towards cleaner and renewable alternatives. Concentrated Solar Power (CSP) technology emerges as a promising solution, offering a carbon-free alternative for power generation. The efficiency and profitability of CSP depend on the Direct Normal Irradiance (DNI) component of solar radiation; hence, accurate DNI forecasting can help optimize CSP plants’ operations and performance. The unpredictable nature of weather phenomena, particularly cloud cover, introduces uncertainty into DNI projections. Existing DNI forecasting models use meteorological factors, which are both challenging to estimate numerically over short prediction windows and expensive to model through data at a sufficiently high spatial and temporal resolution. This research addresses the challenge by presenting a novel approach that formulates DNI prediction as a multi-class classification problem, departing from conventional regression-based methods. The primary objective of this classification framework is to identify optimal periods aligning with specific operational thresholds for CSP plants, contributing to enhanced dispatch optimization strategies. We model the DNI classification problem using four advanced deep neural networks – rectified linear unit (ReLU) networks, 1D residual networks (ResNets), bidirectional long short-term memory (BiLSTM) networks, and transformers – achieving accuracies up to 93.5% without requiring meteorological parameters.

14 SOLAR ENERGY↗
DOE OSTI2011

Development of neural network convection parameterizations for climate and NWP models using Cloud Resolving Model simulations

A novel approach based on the neural network (NN) technique is formulated and used for development of a NN ensemble stochastic convection parameterization for climate models. This fast parameterization is built based on data from Cloud Resolving Model (CRM) simulations initialized with and driven by the TOGA-COARE data available for the 4-month winter season from November 1992 to February 1993. CRM simulated data were averaged and projected onto the GCM space of atmospheric states to implicitly define a stochastic convection parameterization. This parameterization is emulated using an ensemble of neural networks (NNs). The developed NNs are trained and tested. The inherent uncertainty of the stochastic convection parameterization derived following this approach is estimated. The newly developed NN convection parameterization has been tested in a diagnostic mode of NCAR CAM. It produced reasonable and promising climate simulations for the TOGA-COARE winter. This paper is devoted to discussion of the concept, methodology, initial results,and the major challenges of development of NN convection parameterizations for climate and numerical weather prediction (NWR) models.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Geospatial Data Workflow Orchestration and Architecture

In an era characterized by explosive growth in geospatial data, the selection of appropriate technologies for data storage, processing, and orchestration is critical for organizations aiming to maintain competitive advantages. This white paper provides a comprehensive analysis of how Oak Ridge National Laboratory (ORNL) has effectively employed various cloud technologies, including containerized applications, container orchestrators, and workflow orchestrators, to develop robust geospatial data processing solutions. We explore the fundamental concepts behind these technologies and compare multiple deployment models tailored to diverse use cases. Our findings conclude that while Kubernetes has emerged as the preferred platform for truly scalable and fault-tolerant production workflows, the choice of workflow orchestration tool requires careful consideration of team needs, pipeline complexity, and deployment environments. This paper aims to serve as a strategic guide for organizations leveraging geospatial data, articulating the balance between technology choices and practical implementation to enhance workflow efficacy and scalability.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2019

A Plug-Loads Game Changer: Computer Gaming Energy Efficiency without Performance Compromise

Two-thirds of Americans play computer games. Although among the most complex and energy intensive plug loads, gaming has been largely overlooked in energy research and development and policy. Systems used for computer gaming in California consumed 4.1 terawatt-hours/year in 2016 or $700 million in energy bills, with emissions of 1.5 million tons carbon dioxide equivalent allocated 66 percent to consoles, 31 percent to desktop personal computers, 3 percent to laptops, and less than 1 percent to emerging media streaming devices. Key findings include: • Aggregate energy demand places gaming among the top plug loads in California, with gaming representing one-fifth of the state’s total miscellaneous residential energy use. • Market structure changes could substantially affect statewide energy use; energy demand could rise by 114 percent by 2021 under intensified desktop gaming, or fall by 24 percent given a major shift towards consoles coupled with energy efficiency gains. • Unit energy consumption is remarkably varied across gaming platform types: across 26 systems tested, client-side electricity use ranged from 5 to more than 1,200 kWh per year, reflecting equipment choice and usage patterns. • Some emerging technologies and activities are driving energy demand higher, including processor overclocking, cloud-based gaming, higher-resolution connected displays, and virtual reality gaming. • User behavior influences gaming energy use more than technology choice; duty cycle and game choice are particularly strong drivers of demand. • Energy efficiency opportunities are substantial, about 50 percent on a per-system basis for personal computers and 40 percent for consoles if past rates of improvement continue. While simultaneously quantifying efficiency and gaming performance is problematic, evidence suggests that efficiency can be improved while maintaining or improving user experience. Familiar energy policy strategies can help manage gaming energy demand, although mandatory system-level standards are not promising (component-level measures may be).

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2016

High-performance web services for querying gene and variant annotation

Efficient tools for data management and integration are essential for many aspects of high-throughput biology. In particular, annotations of genes and human genetic variants are commonly used but highly fragmented across many resources. Here, we describe MyGene.info and MyVariant.info, high-performance web services for querying gene and variant annotation information. These web services are currently accessed more than three million times permonth. They also demonstrate a generalizable cloud-based model for organizing and querying biological annotation information. MyGene.info and MyVariant.info are provided as high-performance web services, accessible at http://mygene.info and http://myvariant.info. Both are offered free of charge to the research community.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2017

Collisional excitation of CH 2 rotational/fine-structure levels by helium

Accurate determination of the abundance of CH 2 in interstellar media relies on both radiative and collisional rate coefficients. Here, we investigate here the rotational/fine-structure excitation of CH 2 induced by collisions with He. We employ a recoupling technique to generate fine-structure-resolved cross-sections and rate coefficients from close coupling spin-free scattering calculations. The calculations are based on a recent, high-accuracy CH 2 –He potential energy surface computed at the coupled clusters level of theory. The collisional cross-section calculations are performed for all fine-structure transitions among the first 22 and 24 energy levels of ortho- and para-CH 2 , respectively, and for temperatures up to 300 K. As a first application, we simulate the excitation of CH 2 in typical molecular clouds. The excitation temperatures of the CH 2 lines are found to be small at typical densities of molecular clouds, showing that the non-local thermodynamic equilibrium approach has to be used to analyse interstellar spectra. We also found that the fine-structure lines connected with the 404 – 313 and 505 – 414 rotational transitions show possible maser emissions so that they can be easily seen in emission. In conclusion, these calculations show that CH 2 may have to be detected mainly through absorption spectra.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

HAIMOS Ensemble Forecasts for Intra-day and Day- Ahead GHI, DNI and Ramps

The objective of this research is to develop a hybrid physics-based/data-driven forecast model to improve direct normal and global horizontal irradiance (DNI and GHI) prediction for horizons ranging from 1 to 72 hours. Project objectives also address key gaps in state-of-the-art solar forecasting: accurate probabilistic solar forecasts and the forecasting of large irradiance ramps (ramp onset and magnitude). The proposed model ensembles Numerical Weather Prediction (NWP) forecasts, determinist physics-based algorithms, and new-generation cloud cover products (high-resolution rapid refresh satellite images and Large Eddy Simulations). The result is the Hybrid Adaptive Input Model Objective Selection (HAIMOS) ensemble model. HAIMOS blends state of the art machine learning methodologies with physics-based models for cloud cover and cloud optical depth forecasts. The technical activities followed a two-pronged strategy. First, the preprocessing of data, the selection of inputs to the nonlinear approximators, the type of approximator and objective functions, and post-processing ensembling techniques included in HAIMOS were all optimized adaptively to find the best model for a specific goal (reduce DNI/GHI forecast error, improve the prediction of ramp onset, etc.). Second, a large effort was put in improving cloud identification and the forecast of cloud cover and cloud optical depth. To this end, new-generation cloud parametrization products were developed in this work. These include improved algorithms to assist in cloud identification, cloud classification and cloud parametrization from satellite images – three key factors in the accuracy of 1 to 6-hours irradiance forecasts and prediction of ramp onset. Furthermore, we also included cloud information extracted from high resolution rapid refresh satellite images (GOES-16) and Large Eddy Simulations (LES). LES was used to model the atmosphere in detail over locations of interest and produce cloud optical depth forecasts. Once these data streams were validated, they were used as input data to the HAIMOS forecast. The model was developed using data from several climatologically distinct locations with potential for high solar penetration. In the last year of the project, we conducted a validation campaign according to the guidelines stipulated by the Topic Area 1 project as described in the FOA. This effort brings, for the first time, proven machine-learning methodologies for generating state-of-the-art solar forecasts interweaved with detailed physics-based models for cloud detection, and cloud optical depth forecasts. HAIMOS will generate accurate irradiance probabilistic forecast to assist in reducing solar generation prediction error. Globally optimized solar forecast models are more likely to impact solar energy stakeholders. The goal of this project was to increase the state-of-the-art forecast skill from their present values of 10 to 35%. At the end of the project, we achieved between 30% and 50% forecast skill across a wide range of horizons for both GHI and DNI.

14 SOLAR ENERGY↗
DOE OSTI2023

SARS-CoV2 billion-compound docking

Abstract This dataset contains ligand conformations and docking scores for 1.4 billion molecules docked against 6 structural targets from SARS-CoV2, representing 5 unique proteins: MPro, NSP15, PLPro, RDRP, and the Spike protein. Docking was carried out using the AutoDock-GPU platform on the Summit supercomputer and Google Cloud. The docking procedure employed the Solis Wets search method to generate 20 independent ligand binding poses per compound. Each compound geometry was scored using the AutoDock free energy estimate, and rescored using RFScore v3 and DUD-E machine-learned rescoring models. Input protein structures are included, suitable for use by AutoDock-GPU and other docking programs. As the result of an exceptionally large docking campaign, this dataset represents a valuable resource for discovering trends across small molecule and protein binding sites, training AI models, and comparing to inhibitor compounds targeting SARS-CoV-2. The work also gives an example of how to organize and process data from ultra-large docking screens.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2026

Surrogates for Valve-Controlled Pipe Flow: Accelerating Nuclear Reactor Design

Neural surrogate models are developed to replace expensive steady-state RANS CFD simulations for valve-controlled pipe flow in nuclear reactor design. Using parametric CFD data generated with MOOSE Pronghorn across a range of valve geometry and flow conditions, three approaches are compared: a POD-based reduced-order model, a structured UNet on a cylindrical grid, and unstructured models (DeepONet and BiStride MeshGraphNet) on nondimensionalized point clouds. POD achieves the highest accuracy (99%) with fast inference but requires storing all solution snapshots, while the DeepONet and BSMS-GNN both achieve ~89% accuracy at sub-second inference, with the BSMS-GNN offering superior geometric generalizability. These surrogates enable rapid ranking of candidate valve designs and can warm-start CFD solvers to accelerate convergence, supporting agentic design iteration on the Prometheus platform.

42 - ENGINEERING↗
DOE OSTI2020

Panel-Segmentation [SWR-21-18]

Panel-Segmentation contains the scripts for automated metadata extraction of solar PV installations, using satellite imagery coupled with computer vision techniques. In this package, the user can perform the following actions: *Automatically generate a satellite image using a set of lat-long coordinates, and a Google Maps API key. Users would need to set up a Google Cloud account and get a Maps Static API key. Please refer to Setting Up Google Maps Static API Key section for this process. *Perform image segmentation on the satellite image, to locate the solar array(s) in the image on a pixel-by-pixel basis, using an image segmentation model (panel_detection_model.pth). Get classification of the installation (rooftop, ground mounted fixed-tilt or tracking, carport, etc). *Perform azimuth estimation on each solar array cluster in the masked image. *Detect solar panels and get its latitude, longitude, and address within a geographic bounding box through the SOL-Searcher Pipeline. *Detect and calculate hurricane damage on solar installations given pre-hurricane and post-hurricane satellite imagery through the Hurricane Detection Pipeline. *Detect and calculate hail damage on solar installations given satellite imagery through the Hail Detection pipeline. *Convert NOAA MESH (Maximum Estimated Size of Hail) grib2 files into kml or geojson files. *Estimate tilt and azimuth of a solar array by processing USGS LiDAR data for the array’s location.

Edun, Ayobami↗
DOE OSTI2017

Fast Running Urban Dispersion Model for Radiological Dispersal Device (RDD) Releases: Model Description and Validation

Aeolus is an efficient three-dimensional computational fluid dynamics code based on finite volume method developed for predicting transport and dispersion of contaminants in a complex urban area. It solves the time dependent incompressible Navier-Stokes equation on a regular Cartesian staggered grid using a fractional step method. It also solves a scalar transport equation for temperature and using the Boussinesq approximation. The model also includes a Lagrangian dispersion model for predicting the transport and dispersion of atmospheric contaminants. The model can be run in an efficient Reynolds Average Navier-Stokes (RANS) mode with a run time of several minutes, or a more detailed Large Eddy Simulation (LES) mode with run time of hours for a typical simulation. This report describes the model components, including details on the physics models used in the code, as well as several model validation efforts. Aeolus wind and dispersion predictions are compared to field data from the Joint Urban Field Trials 2003 conducted in Oklahoma City (Allwine et al 2004) including both continuous and instantaneous releases. Newly implemented Aeolus capabilities include a decay chain model and an explosive Radiological Dispersal Device (RDD) source term; these capabilities are described. Aeolus predictions using the buoyant explosive RDD source are validated against two experimental data sets: the Green Field explosive cloud rise experiments conducted in Israel (Sharon et al 2012) and the Full-Scale RDD Field Trials conducted in Canada (Green et al 2016).

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

WISP: Watching grid Infrastructure Stealthily through Proxies (Final Technical Report)

The complex interdependencies of cyber systems (sensors and communications), physical grids and associated electricity market operations make protecting electric power grids a significant challenge. The energy sector is constantly under new, targeted, advanced and dangerous cyber-attacks that have the potential to result in the loss of human life. These threats are further exacerbated by our need to modernize the grid. One focus of cyber security research in smart grids is the securing of the SCADA system through advanced intrusion detection systems (IDS) and bad data detection algorithms in state estimation. These methods either require full knowledge of the system topology and parameters or fail to understand the physical behaviors under attack. WISP (Watching grid Infrastructure Stealthily through Proxies) is designed to provide additional protection to the power grid using only publicly available data. In particular, WISP exploits the spatio-temporal nature of the real time locational marginal prices (LMPs), in conjunction with other information such as bids, weather, outages and load data to analyze anomalous power pricing behaviors and then correlate those observations to localize regions of interest and identify potential cyber events. WISP is non-intrusive as the tool is deployed as a service in the Cloud or on premise and provides reliable information to system operators for enhanced situational awareness, without impeding energy delivery functions. The WISP technology comprises three modules: the data-driven anomaly detection core, the vulnerability and risk analysis and the root cause analysis. The data-driven anomaly detection core performs the tasks of feature selection, anomaly detection and attack region localization. The vulnerability and risk analysis module provides system level information of the vulnerable variables and times, assisting the operators in selecting monitoring and protection nodes. The root cause analysis module takes the detection results and identifies potential operational conditions that contribute to the detected anomalies. In Phase I, we have demonstrated the feasibility and effectiveness of WISP. We developed a realistic electricity market simulator capable of generating normal and attack market data under various operational conditions. We developed a series of cyber-attack detection and analysis algorithms and evaluated them under multiple data sources. Finally, we integrated all modules into an end-to-end software, providing functions for data management, data analytics and visualization. Specifically, we have achieved: (i) real-time data acceptance from external utility interfaces with >99% acceptance rate; (ii) high performance anomaly detection algorithms with >98% detection accuracy and <0.1% false alarm rate; and (iii) ultra-low computing delay <50 milliseconds. Additionally, our team developed algorithms to identify the vulnerable variables in electricity market operations and root cause analysis functions to identify major contributors to the price spikes. These ancillary modules are necessary when deploying WISP in real world industry environment. In Phase II, we have demonstrated the effectiveness of WISP software on realistic largescale power systems. We performed red team testing for the Phase I WISP software and identified software vulnerabilities and implemented corresponding mitigation solutions. We adapted the electricity market simulator for the Texas synthetic 2000-bus system and generated datasets for the false data injection attacks. We created database and visualization interfaces for the Texas system and the ISO New England system. We performed software optimization in terms of operation efficiency, computing speed and detection accuracy. Finally, we tested the software on the Texas system and the ISO New England system and evaluated the detection performance. Overall, we achieved above 89% detection rate, below 3% false alarm rate and below 37 seconds of end-to-end detection delay.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2026

Throughput Estimation of Data Transport Networks From Digital Twin Measurements

Digital twins of networked infrastructures, known as Virtual Infrastructure Twins (VITs), are increasingly used for software development, pre-deployment testing, and design space exploration. While VITs avoid the costs and potential disruptions associated with experiments on operational networks, their throughput measurements are typically not sufficiently accurate for performance profiling of wide-area networks that they emulate. Here, machine learning (ML) methods are developed to transform these inaccurate VIT network throughput measurements to closely match in peak and overall profile of those from a physical testbed or production network. First, a micro kernel network reflecting a physical network is utilized to collect one-time measurements on a host to support this ML transformation. Then, a generic multi-modal ML method is developed to learn a map that transforms measurements from subsequent VITs on the same host to match past, current and follow-on testbed and cloud networks. ML generalization equations are derived to establish its correctness and probabilistically guarantee its generalization accuracy. Experimental results are presented for a variety of VIT hosts with target testbed and cloud networks; they include a case study of a four-site science ecosystem wherein inaccurate convex VIT measurement profiles are transformed into accurate concave profiles of target networks.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2012

Application of new simulation algorithms for modeling rf diagnostics of electron clouds

Traveling wave rf diagnostics of electron cloud build-up show promise as a non-destructive technique for measuring plasma density and the efficacy of mitigation techniques. However, it is very difficult to derive an absolute measure of plasma density from experimental measurements for a variety of technical reasons. Detailed numerical simulations are vital in order to understand experimental data, and have successfully modeled build-up. Such simulations are limited in their ability to reproduce experimental data due to the large separation of scales inherent to the problem. Namely, one must resolve both rf frequencies in the GHz range, as well as the plasma modulation frequency of tens of MHz, while running for very long simulations times, on the order of microseconds. The application of new numerical simulation techniques allow us to bridge the simulation scales in this problem and produce spectra that can be directly compared to experiments. The first method is to use a plasma dielectric model to measure plasma-induced phase shifts in the rf wave. The dielectric is modulated at a low frequency, simulating the effects of multiple bunch crossings. This allows simulations to be performed without kinetic particles representing the plasma, which both speeds up the simulations as well as reduces numerical noise from interpolation of particle charge and currents onto the computational grid. Secondly we utilize a port boundary condition model to simultaneously absorb rf at the simulation boundaries, and to launch the rf into the simulation. This method improves the accuracy of simulations by restricting rf frequencies better than adding an external (finite) current source to drive rf, and absorbing layers at the boundaries. We also explore the effects of non-uniform plasma densities on the simulated spectra.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2016

SU-F-P-09: A Global Medical Physics Collaboration for Implementation of Modern Radiotherapy in Botswana

Purpose: The global burden of cancer is considerable, particularly in low and middle-income countries. Massachusetts General Hospital (MGH) and Botswana-Harvard AIDS Institute have partnered with the oncology community and government of Botswana to form BOTSOGO (BOTSwana Oncology Global Outreach) to address the rising burden of cancer in Botswana. Currently, radiation therapy (RT) is only available at a single linear accelerator (LINAC) in Gaborone Private Hospital (GPH). BOTSOGO worked to limit the absence of RT during a LINAC upgrade and ensure a safe transition to modern radiotherapy techniques. Methods: The existing Elekta Precise LINAC was decommissioned in November 2015 and replaced with a new Elekta VERSA-HD with IMRT/VMAT/CBCT capability. Upgraded treatment planning and record-and-verify systems were also installed. Physicists from GPH and MGH collaborated during an intensive on-site visit in Botswana during the commissioning process. Measurements were performed using newly purchased Sun Nuclear equipment. Photon beams were matched with an existing model to minimize the time needed for beam modeling and machine down time. Additional remote peer review was also employed. Independent dosimetry was performed by irradiating OSLDs, which were subsequently analyzed at MGH. Results: Photon beam quality agreed with reference data within 0.2%. Electron beam data agreed with example clinical data within 3%. Absolute dose calibration was performed using both IAEA and AAPM protocols. Absolute dose measurements with OSLDs agreed within 5%. Quentry cloud-based software was installed to facilitate remote review of treatment plans. Patient treatments resumed in February 2016. The time without RT was reduced, therefore likely resulting in reduced patient morbidity/mortality. Conclusion: A global physics collaboration was utilized to commission a modern LINAC in a resource-constrained setting. This can be a useful model in other areas with limited resources. Further use of technology and on-site exchanges will facilitate the introduction of more advanced techniques in Botswana. We acknowledge funding support from the AAPM International Educational Activities Committee and the NCI Federal Share Proton Beam Program Income Grant.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2019

Combined Experimental and Computational Study on the Reaction Dynamics of the 1-Propynyl (CH 3 CC)–1,3-Butadiene (CH 2 CHCHCH 2 ) System and the Formation of Toluene under Single Collision Conditions

In our research, the crossed beams reactions of the 1-propynyl radical (CH 3 CC; X 2 A 1 ) with 1,3-butadiene (CH 2 CHCHCH 2 ; X 1 Ag), 1,3-butadiene- d 6 (CD 2 CDCDCD 2 ; X 1 Ag), 1,3-butadiene- d 4 (CD 2 CHCHCD 2 ; X 1 Ag), and 1,3-butadiene- d 2 (CH 2 CDCDCH 2 ; X 1 Ag) were performed under single collision conditions at collision energies of about 40 kJ mol –1 . The underlying reaction mechanisms were unraveled through the combination of the experimental data with electronic structure calculations at the CCSD(T)-F12/cc-pVTZ-f12//B3LYP/6-311G(d,p) + ZPE(B3LYP/6-311G(d,p) level of theory along with statistical Rice–Ramsperger–Kassel–Marcus (RRKM) calculations. Together, these data suggest the formation of the thermodynamically most stable C 7 H 8 isomer—toluene (C 6 H 5 CH 3 )—via the barrierless addition of 1-propynyl to the 1,3-butadiene terminal carbon atom, forming a low-lying C 7 H 9 intermediate that undergoes multiple isomerization steps resulting in cyclization and ultimately aromatization following hydrogen atom elimination. RRKM calculations predict that the thermodynamically less stable isomers 1,3-heptadien-5-yne, 5-methylene-1,3-cyclohexadiene, and 3-methylene-1-hexen-4-yne are also synthesized. Since the 1-propynyl radical may be present in cold molecular clouds such as TMC-1, this pathway could potentially serve as a carrier of the methyl group incorporating itself into methyl-substituted (poly)acetylenes or aromatic systems such as toluene via overall exoergic reaction mechanisms that are uninhibited by an entrance barrier. Such pathways are a necessary alternative to existing high energy reactions leading to toluene that are formally closed in the cold regions of space and are an important step toward understanding the synthesis of polycyclic aromatic hydrocarbons (PAHs) in space’s harsh extremes.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2020

Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect

High performance multi-GPU computing becomes an inevitable trend due to the ever-increasing demand on computation capability in emerging domains such as deep learning, big data and planet-scale simulations. However, the lack of deep understanding on how modern GPUs can be connected and the real impact of state-of-the-art interconnect technology on multi-GPU application performance become a hurdle. In this paper, we fill the gap by conducting a thorough evaluation on five latest types of modern GPU interconnects: PCIe, NVLink-V1, NVLink-V2, NVLink-SLI and NVSwitch, from six high-end servers and HPC platforms: NVIDIA P100-DGX-1, V100-DGX-1, DGX-2, OLCF’s SummitDev and Summit supercomputers, as well as an SLI-linked system with two NVIDIA Turing RTX-2080 GPUs. Based on the empirical evaluation, we have observed four new types of GPU communication network NUMA effects: three are triggered by NVLink’s topology, connectivity and routing, while one is caused by PCIe chipset design issue. These observations indicate that, for a multi-GPU application running in a multi-GPU node, choosing the right GPU combination can impose considerable impact on GPU communication efficiency, as well as an application’s overall performance. Our evaluation can be leveraged in building practical multi-GPU performance models, which are vital for GPU task allocation, scheduling and migration in a shared environment (e.g., AI cloud and HPC centers), as well as communication-oriented performance tuning.

Performance Evaluation, GPU, Interconnect, NUMA, P↗
DOE OSTI2020

Customizable Scale-Out Key-Value Stores

Enterprise KV stores are often not well suited for HPC applications, and thus cumbersome end-to-end KV design customization is required to meet the needs of modern HPC applications. To this end, in this article we present bespoKV, an adaptive, extensible, and scale-out KV store framework. bespoKV decouples the KV store design into the control plane for distributed management and the data plane for local data store. For the control plane, bespoKVprovides pre-built modules, called controlets, supporting common distributed functionalities (e.g., replication, consistency, and topology) and their various combinations. This decoupling allows bespoKV to take a user-provided single-server KV store, called a datalet, and transparently enables a scalable and fault-tolerant distributed KV store service. The resulting distributed stores are also adaptive to consistency or topology requirement changes and can be easily extended for new types of services. Such specializations enable innovative uses of KV stores in HPC applications, especially for emerging applications that utilize KV-friendly workloads. We evaluate bespoKV in a local testbed as well as in a public cloud settings. Experiments show that bespoKV-enabled distributed KV stores scale horizontally to a large number of nodes, and performs comparably and sometimes 1.2× to 2.6× better than the state-of-the-art systems.

97 MATHEMATICS AND COMPUTING↗