Search NASA⌕ Search

SEARCH · Search NASA

Results for “High Performance Storage Systems”

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 1,009 records · Page 56

The TMI Regenerative Solid Oxide Fuel Cell

Energy storage and production in space requires rugged, reliable hardware which minimizes weight, volume, and maintenance while maximizing power output and usable energy storage. Systems generally consist of photovoltaic solar arrays which operate (during sunlight cycles) to provide system power and regenerate fuel (hydrogen) via water electrolysis and (during dark cycles) fuel cells convert hydrogen into electricity. Common configurations use two separate systems (fuel cell and electrolyzer) in conjunction with photovoltaic cells. Reliability, power to weight and power to volume ratios could be greatly improved if both power production (fuel cells) and power storage (electrolysis) functions can be integrated into a single unit. The solid oxide fuel cell (SOFC) based design integrates fuel cell and electrolyzer functions and potentially simplifies system requirements. The integrated fuel cell/electrolyzer design also utilizes innovative gas storage concepts and operates like a rechargeable 'hydrogen-oxygen battery'. Preliminary research has been completed on improved H2/H20 electrode (SOFC anode/electrolyzer cathode) materials for regenerative fuel cells. Tests have shown improved cell performance in both fuel and electrolysis modes in reversible fuel cell tests. Regenerative fuel cell efficiencies, ratio of power out (fuel cell mode) to power in (electrolyzer mode), improved from 50 percent using conventional electrode materials to over 80 percent. The new materials will allow a single SOFC system to operate as both the electolyzer and fuel cell. Preliminary system designs have also been developed to show the technical feasibility of using the design for space applications requiring high energy storage efficiencies and high specific energy. Small space systems also have potential for dual-use, terrestrial applications.

Cable, Thomas L.↗

Machine Learning Derived Dynamic Operating Reserve Requirements in High-Renewable Power Systems

Accurately forecasting wind and solar power output poses challenges for deeply decarbonized electricity systems. Grid operators must commit resources to provide reserves to ensure reliable operations in the face of forecast errors, a process which can increase fuel consumption and emissions. We apply neural network-based machine learning to expand the usefulness of median point forecast data by creating probabilistic distributions of short-term uncertainty in demand, wind, and solar forecasts that adapt to prevailing grid conditions. Machine learning derived estimates of forecast errors compare favorably to estimates based on incumbent methods. Reserves derived from machine learning are usually smaller than values derived using incumbent methods, which enables fuel savings during most hours. Machine learning reserves are generally larger than incumbent reserves during times of higher forecast error, potentially improving system reliability. Performance is tested using multi-stage production simulation modeling of the California Independent System Operator (CAISO) system. Machine learning reserves provide production cost and greenhouse gas (GHG) emission reductions of approximately 0.3% relative to historical 2019 requirements. Savings in the 2030 timeframe are highly dependent on battery storage capacity. At lower levels of battery capacity, savings of 0.4% from machine learning reserves are shown. Significant quantities of battery storage are expected to be added to meet California's resource adequacy needs and GHG reduction targets. Addition of these batteries saturate reserve needs and results in minimal within-hour balancing costs in 2030.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Energetic Potential for Demand Response in Detached Single Family Homes in Austin, TX

This study builds a generalized tool to quantify the maximum peak load reduction achievable with residential demand response and distributed energy resources (DERs). The tool is demonstrated by using empirical energy usage data from detached single family households within the service territory of Austin Energy - the local municipal utility in Austin, TX. The demand response optimization algorithm shifts energy usage of four controllable, high-consumption residential devices to off-peak hours: HVAC (heating, ventilation, and air conditioning) systems, electric water heaters, electric vehicles, and pool pumps. In addition, rooftop solar generation on individual residences is considered to reduce net load to the utility, and an energy storage system optimization model performing energy arbitrage further reduces peak demand. Results from the aggregate demand response and battery optimization study indicate that peak demand for the 2017 summer peak in Austin Energy could have been reduced by 26 MW or 2.9%. Here, by considering the combined effect of controllable consumer loads, solar generation, and energy storage systems, this study highlights the potential of strategic demand response and energy arbitrage to reduce peak demand at the distribution level.

14 SOLAR ENERGY↗

A Pattern Language for High-Performance Computing Resilience

High-performance computing systems (HPC) provide powerful capabilities for modeling, simulation, and data analytics for a broad class of computational problems. They enable extreme performance of the order of quadrillion floating-point arithmetic calculations per second by aggregating the power of millions of compute, memory, networking and storage components. With the rapidly growing scale and complexity of HPC systems for achieving even greater performance, ensuring their reliable operation in the face of system degradations and failures is a critical challenge. System fault events often lead the scientific applications to produce incorrect results, or may even cause their untimely termination. The sheer number of components in modern extreme-scale HPC systems and the complex interactions and dependencies among the hardware and software components, the applications, and the physical environment makes the design of practical solutions that support fault resilience a complex undertaking. To manage this complexity, we developed a methodology for designing HPC resilience solutions using design patterns. We codified the well-known techniques for handling faults, errors and failures that have been devised, applied and improved upon over the past three decades in the form of design patterns. In this paper, we present a pattern language to enable a structured approach to the development of HPC resilience solutions. The pattern language reveals the relations among the resilience patterns and provides the means to explore alternative techniques for handling a specific fault model that may have different efficiency and complexity characteristics. Using the pattern language enables the design and implementation of comprehensive resilience solutions as a set of interconnected resilience patterns that can be instantiated across layers of the system stack.

Hukerikar, Saurabh↗

A Fusing Switch for Fault Suppression in the SNS High Voltage Converter Modulators

The High Voltage Converter Modulators (HVCMs) at the Spallation Neutron Source (SNS) have operated in excess of a combined 250,000 hours. Performance and reliability improvements to the HVCM are ongoing to increase modulator availability as accelerator system demands increase. There is a relatively large amount of energy storage in the HVCMs, {approx}180 kJ. This energy has the potential to dump into unsuppressed faults, cause damage, and increase the time to repair. The 'fusing switch' concept involves isolation of this stored energy from the location of the most common faults. This paper introduces this concept and its application to the HVCMs.

43 PARTICLE ACCELERATORS↗

Controlled Nucleation and Growth Process of Li 2 S 2 /Li 2 S in Lithium-Sulfur Batteries

Lithium-sulfur battery is a promising next-generation energy storage system because of its potentially three to five times higher energy density than that of traditional lithium ion batteries. However, the dissolution and precipitation of soluble polysulfides during cycling initiate a series of key-chain reactions that significantly shorten battery life. Herein, we demonstrate that through a simple but effective strategy, significantly improved cycling performance is achieved for high sulfur loading electrodes through controlling the nucleation and precipitation of polysulfieds on the electrode surface. More than 400 or 760 stable cycling are successfully displayed in the cells with locked discharge capacity of 625 mAh g -1 or 500 mAh g -1 , respectively. The nucleation and growth process of dissolved polysulfides has been electrochemically altered to confine the thickness of discharge products passivated on the cathode surface, increasing the utilization rate of sulfur while avoiding severe morphology changes on the electrode. More importantly, the exposure of new lithium metal surface to the S-containing electrolyte is also greatly reduced through this strategy, largely minimizing the anode corrosion caused by polysulfides. This work interlocks the electrode morphologies and its evolution with electrochemical interference to modulate cell performances by using Li-S system as a platform, providing different but critical directions for this community.

25 ENERGY STORAGE↗

Best Practices for Operation and Maintenance of Photovoltaic and Energy Storage Systems; 3rd Edition

The goal of this guide is to reduce the cost and improve the effectiveness of operations and maintenance (O&M) for photovoltaic (PV) systems and combined PV and energy storage systems. Reported O&M costs vary widely based on the requirements of the system and the nature of the O&M contract, but a more standardized approach to planning and delivering O&M has the potential to both decrease costs and make those costs more predictable over time. This guide encourages adoption of best practices to reduce the cost of O&M and improve the performance of fielded systems, but it also enhances financing of new projects by making cost more predictable and mitigating performance risk. This guide also includes technical improvement opportunities in the design of systems and in specification of equipment because high-quality system deployment improves lifetime project performance and energy production while reducing, or at least optimizing, costs to deliver an O&M program. This guide provides recommendations that increase the effectiveness of O&M services; reduce O&M costs, improve solar asset transparency for investors and rating agencies; provide an industry framework for quality management; and reduce transaction costs.

14 SOLAR ENERGY↗

Software-defined network for end-to-end networked science at the exascale

Domain science applications and workflow processes are currently forced to view the network as an opaque infrastructure into which they inject data and hope that it emerges at the destination with an acceptable Quality of Experience. There is little ability for applications to interact with the network to exchange information, negotiate performance parameters, discover expected performance metrics, or receive status/troubleshooting information in real time. The work presented here is motivated by a vision for a new smart network and smart application ecosystem that will provide a more deterministic and interactive environment for domain science workflows. The Software-Defined Network for End-to-end Networked Science at Exascale (SENSE) system includes a model-based architecture, implementation, and deployment which enables automated end- to-end network service instantiation across administrative domains. An intent based interface allows applications to express their high-level service requirements, an intelligent orchestrator and resource control systems allow for custom tailoring of scalability and real-time responsiveness based on individual application and infrastructure operator requirements. This allows the science applications to manage the network as a first-class schedulable resource as is the current practice for instruments, compute, and storage systems. Deployment and experiments on production networks and testbeds have validated SENSE functions and performance. Emulation based testing verified the scalability needed to support research and education infrastructures. Key contributions of this work include an architecture definition, reference implementation, and deployment. This provides the basis for further innovation of smart network services to accelerate scientific discovery in the era of big data, cloud computing, machine learning and artificial intelligence.

47 OTHER INSTRUMENTATION↗

Software-Defined Network for End-to-end Networked Science at the Exascale

Domain science applications and workflow processes are currently forced to view the network as an opaque infrastructure into which they inject data and hope that it emerges at the destination with an acceptable Quality of Experience. There is little ability for applications to interact with the network to exchange information, negotiate performance parameters, discover expected performance metrics, or receive status/troubleshooting information in real time. The work we presen here is motivated by a vision for a new smart network and smart application ecosystem that will provide a more deterministic and interactive environment for domain science workflows. The Software-Defined Network for End-to-end Networked Science at Exascale (SENSE) system includes a model-based architecture, implementation, and deployment which enables automated end-to-end network service instantiation across administrative domains. An intent based interface allows applications to express their high-level service requirements, an intelligent orchestrator and resource control systems allow for custom tailoring of scalability and real-time responsiveness based on individual application and infrastructure operator requirements. This allows the science applications to manage the network as a first-class schedulable resource as is the current practice for instruments, compute, and storage systems. Deployment and experiments on production networks and testbeds have validated SENSE functions and performance. Emulation based testing verified the scalability needed to support research and education infrastructures. Key contributions of this work include an architecture definition, reference implementation, and deployment. This provides the basis for further innovation of smart network services to accelerate scientific discovery in the era of big data, cloud computing, machine learning and artificial intelligence.

97 MATHEMATICS AND COMPUTING↗

Accelerating Battery Simulations by Using High Performance Computing and Opportunities with Machine Learning

The US Department of Energy (DOE) estimates that the battery pack cost of $60 per kilowatt-hour while increasing the driving range to over 300 miles and vehicle charging under 15 min or less would enable mass penetration of electric vehicles in the USA by 2030. These projections are based on the currently available high-density cell chemistry combined with a system level design and optimized electrodes. Electrodes for current state-of-the-art lithium-ion cell technology are fabricated from electrode slurry comprising of active material, polymeric binder, and conductive diluent such as carbon black that are coated on metal current collectors such as copper and aluminum. Given these challenging requirements for development of electrical energy storage devices for future transportation needs, a predictive simulation capability which can accelerate design by considering performance and safety implications of different geometry, materials, and chemistry choices is required. In this chapter, we discuss our state-of-the-art three-dimensional modeling framework, providing examples on battery performance and safety simulations. We also present approaches to use machine learning in the context of large datasets that are becoming available.

Allu, Srikanth↗

Characterizing output bottlenecks in a supercomputer

Supercomputer I/O loads are often dominated by writes. HPC (High Performance Computing) file systems are designed to absorb these bursty outputs at high bandwidth through massive parallelism. However, the delivered write bandwidth often falls well below the peak. This paper characterizes the data absorption behavior of a center-wide shared Lustre parallel file system on the Jaguar supercomputer. We use a statistical methodology to address the challenges of accurately measuring a shared machine under production load and to obtain the distribution of bandwidth across samples of compute nodes, storage targets, and time intervals. We observe and quantify limitations from competing traffic, contention on storage servers and I/O routers, concurrency limitations in the client compute node operating systems, and the impact of variance (stragglers) on coupled output such as striping. We then examine the implications of our results for application performance and the design of I/O middleware systems on shared supercomputers.

Xie, Bing↗

Characterizing output bottlenecks in a supercomputer

Supercomputer I/O loads are often dominated by writes. HPC (High Performance Computing) file systems are designed to absorb these bursty outputs at high bandwidth through massive parallelism. However, the delivered write bandwidth often falls well below the peak. This paper characterizes the data absorption behavior of a center-wide shared Lustre parallel file system on the Jaguar supercomputer. We use a statistical methodology to address the challenges of accurately measuring a shared machine under production load and to obtain the distribution of bandwidth across samples of compute nodes, storage targets, and time intervals. We observe and quantify limitations from competing traffic, contention on storage servers and I/O routers, concurrency limitations in the client compute node operating systems, and the impact of variance (stragglers) on coupled output such as striping. We then examine the implications of our results for application performance and the design of I/O middleware systems on shared supercomputers.

Xie, Bing↗

Demonstration of High-Temperature Calcium-Based Thermochemical Energy Storage System for use with Concentrating Solar Power Facilities

Thermochemical energy storage (TCES) systems offer several advantages over sensible heat storage via molten salts for concentrating solar power (CSP) including: energy density, cost, indefinite storage time, and no risk of freezing. SR’s TCES system utilizes a CaO-based carbonation reaction at high temperature in a fixed bed heat exchanger reactor. The SunShot APOLLO project focuses on the validation of the calcium oxide sorbent for capacity, durability, system economics, and demonstration of the system at a small pilot scale, including system control and operation in simulated on-sun conditions, discharging at 720°C. Under phase I SR completed evaluation of the sorbent loaded in a single-channel heat exchanger reactor. The major accomplishments of phase II are: Design, fabrication and commissioning of a 4 kWhth pilot plant (50x capacity increase from previous bench scale) with a closed-loop CO 2 system, and a commercially relevant heat exchange reactor. The heat exchange reactor has been designed with state-of-the-art modeling tools to ensure 1) sufficient lifetime of the Incoloy body and 2) performance of the packed bed reactor will be adequate to meet the milestones. The rest of the pilot system has been designed with off-the shelf components to maximize reliability and cost effectiveness. The system is highly instrumented and capable of time resolved, independent measurements of the heat and mass balances respectively. Major developments in the technoeconomic analysis (TEA) include 1) reduction of the cost of membrane gas storage for the near-term and future systems and 2) integration with liquid metal receivers and particle receivers for higher decarbonation temperatures and pressures which reduced the cost of gas compression and increased system efficiency. The result are three scenarios with costs bases of 21.9 dollars, 14.3 dollars, and 10.9 dollars/kWh th corresponding to a real LCOE of 59, 57, 56 dollars/MW he , respectively. The basic operations of the energy storage system were demonstrated for carbonation (discharging) and decarbonation (charging) over short time periods. Excellent control of the CO 2 mass balance across the reactor was achieved with mass flow controllers and the basic operation of the level control of the CO 2 gas bladder was achieved. The system was apparently free from leaks and operated continuously for 18 days with no mechanical failures, thus meeting milestone 2.5.5. The major shortcomings of the system had to do with the controls which prevented us from deep cycling and or reaching high cycles. We were unable to decarbonate (charge) the system for more than a few minutes at a time, which was much shorter than the design target of 10 hrs. This issue could not be resolved due to facility availability and contract closure. One major finding is that CaO-based TCES is most cost efficient when charged with HTF (heat transfer fluid) temperatures >850°C to reach the round-trip exergy efficiency and system cost targets which are higher temperatures than those considered for Gen 3 CSP. Many of the balance of plant components do not exist today, however we are encouraged by our findings that in general with more work we are able to reduce the capital cost, for example with the membrane gas storage, which is an innovative component in itself. The CaO-based TCES system has potential to meet the latest SunShot 2030 goals of LCOE at 50 $/MW he with capital costs of <11/kWh th .

14 SOLAR ENERGY↗

Demonstration of High-Temperature Calcium-Based Thermochemical Energy Storage System for use with Concentrating Solar Power Facilities

Thermochemical energy storage (TCES) systems offer several advantages over sensible heat storage via molten salts for concentrating solar power (CSP) including: energy density, cost, indefinite storage time, and no risk of freezing. SR’s TCES system utilizes a CaO-based carbonation reaction at high temperature in a fixed bed heat exchanger reactor. The SunShot APOLLO project focuses on the validation of the calcium oxide sorbent for capacity, durability, system economics, and demonstration of the system at a small pilot scale, including system control and operation in simulated on-sun conditions, discharging at 720°C. Under phase I SR completed evaluation of the sorbent loaded in a single-channel heat exchanger reactor. The major accomplishments of phase II are: Design, fabrication and commissioning of a 4 kWhth pilot plant (50x capacity increase from previous bench scale) with a closed-loop CO 2 system, and a commercially relevant heat exchange reactor. The heat exchange reactor has been designed with state-of-the-art modeling tools to ensure 1) sufficient lifetime of the Incoloy body and 2) performance of the packed bed reactor will be adequate to meet the milestones. The rest of the pilot system has been designed with off-the shelf components to maximize reliability and cost effectiveness. The system is highly instrumented and capable of time resolved, independent measurements of the heat and mass balances respectively. Major developments in the technoeconomic analysis (TEA) include 1) reduction of the cost of membrane gas storage for the near-term and future systems and 2) integration with liquid metal receivers and particle receivers for higher decarbonation temperatures and pressures which reduced the cost of gas compression and increased system efficiency. The result are three scenarios with costs bases of 21.9 dollars, 14.3 dollars, and 10.9 dollars/kWhth corresponding to a real LCOE of 59, 57, 56 dollars/MWhe, respectively. The basic operations of the energy storage system were demonstrated for carbonation (discharging) and decarbonation (charging) over short time periods. Excellent control of the CO 2 mass balance across the reactor was achieved with mass flow controllers and the basic operation of the level control of the CO 2 gas bladder was achieved. The system was apparently free from leaks and operated continuously for 18 days with no mechanical failures, thus meeting milestone 2.5.5. The major shortcomings of the system had to do with the controls which prevented us from deep cycling and or reaching high cycles. We were unable to decarbonate (charge) the system for more than a few minutes at a time, which was much shorter than the design target of 10 hrs. This issue could not be resolved due to facility availability and contract closure. One major finding is that CaO-based TCES is most cost efficient when charged with HTF (heat transfer fluid) temperatures >850°C to reach the round-trip exergy efficiency and system cost targets which are higher temperatures than those considered for Gen 3 CSP. Many of the balance of plant components do not exist today, however we are encouraged by our findings that in general with more work we are able to reduce the capital cost, for example with the membrane gas storage, which is an innovative component in itself. The CaO-based TCES system has potential to meet the latest SunShot 2030 goals of LCOE at 50 dollars/MWhe with capital costs of <11/kWhth.

14 SOLAR ENERGY↗

Advanced battery modeling for interfacial phenomena and optimal charging

Lithium ion batteries are one of the most promising energy storage systems for portable devices, transportation, and renewable grids. To meet the increasing requirements of these applications, higher energy density and areal capacity, long cycle life, fast charging rate and enhanced safety for lithium ion battery (LIBs) are urgently needed. To solve these challenges, the relevant physics at different length scale need to be understood. However, experimental study is time consuming and limited in small scale’s study. Modeling techniques provide us powerful tools to get a deep understanding of the relevant physics and find optimal solutions. This work focuses on studying the mechanism in advanced battery engineering techniques and developing a new charging algorithm by model-based optimization. The research topics are divided into six topics and each topic is reported as a form of journal publication. Paper Ⅰ provides a new aspect of how ALD coating improves the lithium ion diffusion at electrode particles. Paper Ⅱ explains the mechanisms by which 3D electrodes enhance battery performance and reveals guidelines for optimized 3D electrode designs by a 3D electrochemical-mechanical battery model. Paper Ⅲ investigates the electrolyte concentration impact on SEI layer growth and Li plating, especially under high charge rates. Paper Ⅳ proposes an optimized charging protocol for fast charging for reducing the charging time with minimal degradation. Paper Ⅴ reports a comprehensive degradation model for degradation estimation and life predication of energy storage system (ESS). Paper Ⅵ is a study of temperature-dependent state of charge (SOC) estimation for battery pack.

25 ENERGY STORAGE↗

Influence of Metal Identity on Light-Induced Switchable Adsorption in Azobenzene-Based Metal–Organic Frameworks

Energy-efficient capture and release of small gas molecules, particularly carbon dioxide (CO 2 ) and methane (CH 4 ), are of significant interest in academia and industry. Porous materials such as metal–organic frameworks (MOFs) have been extensively studied, as their ultrahigh porosities and tunability enable significant amounts of gas to be adsorbed while also allowing specific applications to be targeted. However, because of the microporous nature of MOFs, the gas adsorption performance is dominated by high uptake capacity at low pressures, limiting their application. Hence, methods involving stimuli-responsive materials, particularly light-induced switchable adsorption (LISA), offer a unique alternative to thermal methods. Here, we report the mechanism of a well-known LISA system, the azobenzene-based material PCN-250, for CO 2 and CH 4 adsorption. There is a noticeable difference in the LISA effect dependent on the metal cluster involved, with the most significant being PCN-250-Al, where the adsorption can change by 83.1% CH 4 and 56.1% CO 2 at 298 K and 1 bar and inducing volumetric storage changes of 36.2 and 33.9 cm 3 /cm 3 at 298 K between 5 and 85 bar (CH 4 ) and 2 and 9 bar (CO 2 ), respectively. Using UV light in both single-crystal X-ray diffraction and gas adsorption testing, we show that upon photoirradiation, the framework undergoes a “localized heating” phenomenon comparable to an increase of 130 K for PCN-250-Fe and improves the working capacity. Furthermore, this process functions because of the constrained nature of the ligand, preventing the typical trans-to-cis isomerization observed in free azobenzene. In addition, we observed that the degree of localized heating is highly dependent on the metal cluster involved, with the series of isostructural PCN-250 systems showing variable performance based upon the degree of interaction between the ligand and the metal center.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

NRC Technical Research Program to Evaluate Extended Storage and Transportation of Spent Nuclear Fuel - 12547

Any new direction proposed for the back-end of spent nuclear fuel (SNF) cycle will require storage of SNF beyond the current licensing periods. The Nuclear Regulatory Commission (NRC) has established a technical research program to determine if any changes in the 10 CFR part 71, and 72 requirements, and associated guidance might be necessary to regulate the safety of anticipated extended storage, and subsequent transport of SNF. This three part program of: 1) analysis of knowledge gaps in the potential degradation of materials, 2) short-term research and modeling, and 3) long-term demonstration of systems, will allow the NRC to make informed regulatory changes, and determine when and if additional monitoring and inspection of the systems is necessary. The NRC has started a research program to obtain data necessary to determine if the current regulatory guidance is sufficient if interim dry storage has to be extended beyond the currently approved licensing periods. The three-phased approach consists of: - the identification and prioritization of potential degradation of the components related to the safe operation of a dry cask storage system, - short-term research to determine if the initial analysis was correct, and - a long-term prototypic demonstration project to confirm the models and results obtained in the short-term research. The gap analysis has identified issues with the SCC of the stainless steel canisters, and SNF behavior. Issues impacting the SNF and canister internal performance such as high and low temperature distributions, and drying have also been identified. Research to evaluate these issues is underway. Evaluations have been conducted to determine the relative values that various types of long-term demonstration projects might provide. These projects or follow-on work is expected to continue over the next five years. (authors)

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

Hvac: Removing I/O Bottleneck for Large-Scale Deep Learning Applications

Scientific communities are increasingly adopting deep learning (DL) models in their applications to accelerate scientific discovery processes. However, with rapid growth in the computing capabilities of HPC supercomputers, large-scale DL applications have to spend a significant portion of training time performing I/O to a parallel storage system. Previous research works have investigated optimization techniques such as prefetching and caching. Unfortunately, there exist non-trivial challenges to adopting the existing solutions on HPC supercomputers for large-scale DL training applications, which include non-performance and/or failures at extreme scale, lack of portability and generality in design, complex deployment methodology, and being limited to a specific application or dataset. To address these challenges, we propose High-Velocity AI Cache (HVAC), a distributed read-cache layer that targets and fully exploits the node-local storage or near node-local storage technology. HVAC seamlessly accelerates read I/O by aggregating node-local or near node-local storage, avoiding metadata lookups and file locking while preserving portability in the application code. We deploy and evaluate HVAC on 1,024 nodes (with over 6000 NVIDIA V100 GPUS) of the Summit supercomputer. In particular, we evaluate the scalability, efficiency, accuracy, and load distribution of HVAC compared to GPFS and XFS-on-NVMe. With four different DL applications, we observe an average 25 % performance improvement atop GPFS and 9% drop against XFS-on-NVMe, which scale linearly and are considered the performance upper bound. We envision HVAC as an important caching library for upcoming HPC supercomputers such as Frontier.

Khan, Awais↗