Search NASA⌕ Search

SEARCH · Search NASA

Results for “online learning algorithm”

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 145 records · Page 8

A reporting format for field measurements of soil respiration

Field observations of the soil-to-atmosphere CO2 flux–soil respiration, RS–are a prime example of ‘long tail’ data that historically have had neither centralized databases nor an agreed-upon reporting format. This has hindered scientific transparency, analytical reproducibility, and novel syntheses with respect to this globally-important component of the carbon cycle. Here we propose a new data and metadata reporting format for RS data, based on engagement with a wide range of researchers in the field as well as expert advisory panels. Our goal was a reporting format that would be relevant and useful for synthesis activities, and optimizing data discoverability and usability while not placing an undue burden on data contributors. We describe previous RS data collection efforts, lessons learned from related databases and data-oriented networks (e.g. FLUXNET) in earth and ecological sciences, and the process of community consultation. The proposed reporting format focuses on chamber-level data and metadata, specifying measurement conditions and, for a given measurement period defined by beginning and ending timestamps, a mean RS flux (or CO2 concentration) and associated ancillary measurements. Fundamentally, this format aims to enable findable, accessible, interoperable, and reusable data, while providing ‘future-proofing’ capabilities to support reanalyses using as yet unknown algorithms or approaches. Finally, this proposed RS reporting format is available online, and is intended to be a dynamic document, subject to further community feedback and/or change in the future.

Bond-Lamberty, Benjamin↗

Grid Optimization with Solar (GO-Solar) Experiences With: Data-Driven and Machine Learning Approaches for High-Pen PV Grids

Provides an overview of the innovations and challenges from the algorithm development portion of the GO-Solar project, including plain language introductions to matrix-completion based state estimation, multi-kernel learning based state forecasting, 'slow-scale' voltage-load sensitivity matrix linearized optimal power flow, and 'fast-scale' OPF plan following using on-line Provides an overview of the innovations and challenges from the algorithm development portion of the GO-Solar project, including plain language introductions to matrix-completion based state estimation, multi-kernel learning based state forecasting, 'slow-scale' voltage-load sensitivity matrix linearized optimal power flow, and 'fast-scale' OPF plan following using on-line multi-objective state.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Learning-Based Load Control to Support Resilient Networked Microgrid Operations

Microgrids have proven to be an effective option for increasing the resiliency of critical end-use loads during extreme events. Building on past operational experiences, some microgrid operators are examining the potential to network microgrids to further improve resiliency. However, the frequency deviations experienced on isolated microgrids during transient events, such as switching operations, step increases in load, and loss of generation, are significantly larger than those typically seen on bulk transmission systems. The larger frequency deviations can cause a loss of inverter-connected assets, resulting in a loss of power to critical end-use loads. This paper presents a method of mitigating the impact of transient events by engaging end-use loads using Grid-Friendly Appliance TM (GFA) controllers. An online, i.e., real-time, device-level algorithm is presented, which adjusts individual GFA controller frequency set-points based on the operational characteristics of each end-use load, and on the changing grid dynamic characteristics. The presented method improves the dynamic stability of the networked microgrid operations while minimizing the interruptions to end-use loads. The presented work is validated with dynamic simulations using a modified version of the IEEE 123-node test system with three microgrids, using the GridLAB-D TM simulation environment.

Radhakrishnan, Nikitha↗

Real-Time Health Monitoring for Gas Turbine Components Using Online Learning and High-Dimensional Data

Capital-intensive turbomachinery, such as gas turbines and combined cycle plants, are constantly being monitored for performance anomalies, faults, and physical degradation. Although these power-generating assets are equipped with hundreds of sensors, existing monitoring tools can only handle moderate-sized data. As a result, only a handful of aggregate metrics are used to monitor machine health. At the same time, developing advanced tools suitable for large datasets have been restricted by the lack of appropriate data. The objective of this proposal was to demonstrate a Big Data analytics framework for fault detection and diagnosis in gas turbine applications. We develop a predictive analytics framework methodology guided by these experimental data, industrial data from our collaborators, and physics-based models with engineering domain knowledge. Our analytics framework consists of four key components (1) a data curation process that addresses data storage, data quality assessments, and integrity checks, (2) a feature engineering component that utilizes statistical methods and transformation algorithms guided by physics-based models to extract high-fidelity fault features that can be leveraged for fault detection and classifying fault severities, (3) a Machine Learning-based fault detection and diagnostics algorithms for detecting operational and hardware faults in the combustion and the turbines section. We utilize two industry-class gas turbine component test rigs to generate first of its kind data for critical gas turbine faults with varying severity levels. Advanced gas turbine test facilities will be interrogated using state-of-the-art instrumentation techniques to build fault signatures and data trends for key combustor and turbine faults. Data generated from a combustor test rig (Georgia Tech) and a turbine test rig (Penn State) during both normal operation and with seeded faults serve as the basis for the Big Data sets. The test conditions in the two test facilities include common, critical events that occur in the operation. Utilizing the combustor test rig, we examine two common combustor faults: lean blowout and centerbody degradation. For the turbine section we develop analytic models for monitoring cooling faults in the gas turbine

03 NATURAL GAS↗

Real-Time Health Monitoring for Gas Turbine Components Using Online Learning and High-Dimensional Data (Final Report)

Capital-intensive turbomachinery, such as gas turbines and combined cycle plants, are constantly being monitored for performance anomalies, faults, and physical degradation. Although these power-generating assets are equipped with hundreds of sensors, existing monitoring tools can only handle moderate-sized data. As a result, only a handful of aggregate metrics are used to monitor machine health. At the same time, developing advanced tools suitable for large datasets have been restricted by the lack of appropriate data. The objective of this proposal was to demonstrate a Big Data analytics framework for fault detection and diagnosis in gas turbine applications. We develop a predictive analytics framework methodology guided by these experimental data, industrial data from our collaborators, and physics-based models with engineering domain knowledge. Our analytics framework consists of four key components: (1) a data curation process that addresses data storage, data quality assessments, and integrity checks, (2) a feature engineering component that utilizes statistical methods and transformation algorithms guided by physics-based models to extract high-fidelity fault features that can be leveraged for fault detection and classifying fault severities, (3) a Machine Learning-based fault detection and diagnostics algorithms for detecting operational and hardware faults in the combustion and the turbines section. We utilize two industry-class gas turbine component test rigs to generate first of its kind data for critical gas turbine faults with varying severity levels. Advanced gas turbine test facilities will be interrogated using state-of-the-art instrumentation techniques to build fault signatures and data trends for key combustor and turbine faults. Data generated from a combustor test rig (Georgia Tech) and a turbine test rig (Penn State) during both normal operation and with seeded faults serve as the basis for the Big Data sets. The test conditions in the two test facilities include common, critical events that occur in the operation. Utilizing the combustor test rig, we examine two common combustor faults: lean blowout and centerbody degradation. For the turbine section we develop analytic models for monitoring cooling faults in the gas turbine.

20 FOSSIL-FUELED POWER PLANTS↗

Online data-driven changepoint detection for high-dimensional dynamical systems

In this study, the detection of anomalies or transitions in complex dynamical systems is of critical importance to various applications. In this study, we propose the use of machine learning to detect changepoints for high-dimensional dynamical systems. Here, changepoints indicate instances in time when the underlying dynamical system has a fundamentally different characteristic—which may be due to a change in the model parameters or due to intermittent phenomena arising from the same model. We propose two complementary approaches to achieve this, with the first devised using arguments from probabilistic unsupervised learning and the latter devised using supervised deep learning. To accelerate the deployment of transition detection algorithms in high-dimensional dynamical systems, we introduce dimensionality reduction techniques. Our experiments demonstrate that transitions can be detected efficiently, in real-time, for the two-dimensional forced Kolmogorov flow and the Rössler dynamical system, which are characterized by anomalous regimes in phase space where dynamics are perturbed off the attractor at potentially uneven intervals. Finally, we also demonstrate how variations in the frequency of detected changepoints may be utilized to detect a significant modification to the underlying model parameters by utilizing the Lorenz-63 dynamical system.

97 MATHEMATICS AND COMPUTING↗

Signal Decomposition for Intrusion Detection in Reliability Assessment in Cyber Resilience (Summary Report)

The complexity of assuring cyber resilience for physical process interactions in connected systems such as energy grids increases dramatically as the coupling between processes becomes more direct and responsive. An example of this growing complexity is provided by Integrated Energy Systems (IES), in which various processes such as nuclear heat generation and commodity production are being directly coupled for increased responsiveness to highly variable signals such as market pricing or electricity demand. As such, the potential attack surface of the coupled processes is larger than the two processes independently. Securing these complex systems requires two-fold monitoring: cybersecure monitoring for potential malicious incursion, and physics monitoring for system tampering. Physics monitoring includes analyzing the behavior of the signals within the system for anomalous behavior. This analysis has been shown to be insufficient if approached by only data-driven machine learning and artificial intelligence (MLAI) techniques or only low-level model comparison. Previous efforts at Purdue University suggested combining high-fidelity models with MLAI algorithms as a basis for a software tool for detecting anomalies in physical processes. This work built on that suggestion, developing an advanced library for signal decomposition and analysis using both MLAI and high-fidelity physics algorithms for greatly improved anomaly detection, especially false data injection. This software can be used as part of a secure imbedded intelligence (SEI) system designed under Consequence-driven Cyber-informed Engineering (CCE) for complex coupled systems. This library established a foundation for online and posteriori analysis of digital signals for the purpose of detecting potential malicious tampering in digital signals representing physical processes. Demonstrations carried out throughout the development highlight the effective use of characterization algorithms to detect signal perturbations, particularly triangle attack-style perturbations, in three wide-ranging applications: seismic monitoring, nuclear thermal hydraulics system simulation, and custom manufacturing.

97 MATHEMATICS AND COMPUTING↗

A Magnetoencephalographic/Encephalographic (MEG/EEG) Brain-Computer Interface Driver for Interactive iOS Mobile Videogame Applications Utilizing the Hadoop Ecosystem, MongoDB, and Cassandra NoSQL Databases

In Phase I, we collected data on five subjects yielding over 90% positive performance in Magnetoencephalographic (MEG) mid-and post-movement activity. In addition, a driver was developed that substituted the actions of the Brain Computer Interface (BCI) as mouse button presses for real-time use in visual simulations. The process was interfaced to a flight visualization demonstration utilizing left or right brainwave thought movement, the user experiences, the aircraft turning in the chosen direction, or on iOS Mobile Warfighter Videogame application. The BCI’s data analytics of a subject’s MEG brain waves and flight visualization performance videogame analytics were stored and analyzed using the Hadoop Ecosystem as a quick retrieval data warehouse. In Phase II portion of the project involves the Emotiv Encephalographic (EEG) Wireless Brain–Computer interfaces (BCIs) allow for people to establish a novel communication channel between the human brain and a machine, in this case, an iOS Mobile Application(s). The EEG BCI utilizes advanced and novel machine learning algorithms, as well as the Spark Directed Acyclic Graph (DAG), Cassandra NoSQL database environment, and also the competitor NoSQL MongoDB database for housing BCI analytics of subject’s response and users’ intent illustrated for both MEG/EEG brainwave signal acquisition. The wireless EEG signals that were acquired from the OpenVibe and the Emotiv EPOC headset can be connected via Bluetooth to an iPhone utilizing a thin Client architecture. The use of NoSQL databases were chosen because of its schema-less architecture and Map Reduce computational paradigm algorithm for housing a user’s brain signals from each referencing sensor. Thus, in the near future, if multiple users are playing on an online network connection and an MEG/EEG sensor fails, or if the connection is lost from the smartphone and the webserver due to low battery power or failed data transmission, it will not nullify the NoSQL document-oriented (MongoDB) or column-oriented Cassandra databases. Additionally, NoSQL databases have fast querying and indexing methodologies, which are perfect for online game analytics and technology. In Phase II, we collected data on five MEG subjects, yielding over 90% positive performance on iOS Mobile Applications with Objective-C and C++, however on EEG signals utilized on three subjects with the Emotiv wireless headsets and (n < 10) subjects from the OpenVibe EEG database the Variational Bayesian Factor Analysis Algorithm (VBFA) yielded below 60% performance and we are currently pursuing extending the VBFA algorithm to work in the time-frequency domain referred to as VBFA-TF to enhance EEG performance in the near future. The novel usage of NoSQL databases, Cassandra and MongoDB, were the primary main enhancements of the BCI Phase II MEG/EEG brain signal data acquisition, queries, and rapid analytics, with MapReduce and Spark DAG demonstrating future implications for next generation biometric MEG/EEG NoSQL databases.

McClay, Wilbert↗

A mapped dataset of surface ocean acidification indicators in large marine ecosystems of the United States

Mapped monthly data products of surface ocean acidification indicators from 1998 to 2022 on a 0.25° by 0.25° spatial grid have been developed for eleven U.S. large marine ecosystems (LMEs). The data products were constructed using observations from the Surface Ocean CO 2 Atlas, co-located surface ocean properties, and two types of machine learning algorithms: Gaussian mixture models to organize LMEs into clusters of similar environmental variability and random forest regressions (RFRs) that were trained and applied within each cluster to spatiotemporally interpolate the observational data. The data products, called RFR-LMEs, have been averaged into regional timeseries to summarize the status of ocean acidification in U.S. coastal waters, showing a domain-wide carbon dioxide partial pressure increase of 1.4 ± 0.4 μatm yr -1 and pH decrease of 0.0014 ± 0.0004 yr -1 . RFR-LMEs have been evaluated via comparisons to discrete shipboard data, fixed timeseries, and other mapped surface ocean carbon chemistry data products. Regionally averaged timeseries of RFR-LME indicators are provided online through the NOAA National Marine Ecosystem Status web portal.

54 ENVIRONMENTAL SCIENCES↗

Kernelized approaches to streaming compression of scientific data

In this paper three algorithms are developed for the streaming compression of scientific data. The algorithms presented are reliant on the theory of vector-valued reproducing kernel Hilbert spaces and operator valued kernel. Further, the scientific data is modeled as a snapshot of time dependent vector field F(x, t) over a manifold M and the recovery of the data is framed as a learning problem. These processes are then appropriately modified and ana lyzed for the streaming scenario in which data is generated without the ability to revisit past entries.

97 MATHEMATICS AND COMPUTING↗

A simplified regression building thermal modelling method for detached two-floor house in U.S.

The residential building sector accounts for approximately 37% of total U.S. electricity consumption. Within the residential building sector, heating and cooling is the main target for peak load shifting/reduction since it is the largest contributor to peak demand. In fact, the flexibility of residential HVAC loads can provide continuous variation of demand to provide grid services by varying their demand over a baseline. The performance of HVAC load control to provide grid services relies heavily on the accuracy of indoor air temperature or cooling/heating demand predictions and therefore the quality of building model. Besides forward models, popular building models are data-driven models which can be broken down into two categories: gray-box model, e.g. Resistance-Capacitance (RC) model and black-box model. RC model, also called lumped capacitance or network model, which is constituted with electrical analogue pattern with resistance (R) and capacitance (C). In general, RC models require considerable computation burden and long periods of data to train limited number of model coefficients. Black-box models have gained increasing interest due to their capability in analyzing large-scale data and flexibility in practical applications. But, the data-mining based (machine learning algorithms/techniques based) models tend to have invisible model structures which poses a problem when trying to use the model for optimal control or model predictive control of the HVAC system. Hence, there is a continuing need for efficient online system identification techniques, which can provide explicit parameters for the model. Traditional regression models fit well for this specific purpose. This paper presents an innovative way to predict average indoor temperature in separate floors of typical detached residential house. A rolling horizon linear regression model, which includes online adaptive correction component, is proposed to predict the temperature difference between downstairs and upstairs. A RC model is used to predict the overall mean indoor air temperature. Since the adaptive algorithm needs to be implemented online, a less computation-demanding polynomial fitting algorithm is adopted. This kind of fitting problem can be cast as linear regression problem with multiple variables, parameters of which can be efficiently obtained by well-known gradient descent method.The validation is conducted by comparing the predicted results with the results from data-mining based models as well as measured data from a real typical detached two-floor house. The results show that the developed method has satisfactory performance in predicting the building indoor temperature in 1st and 2nd floors.

Cui, Borui↗

Deep Learning Image Segmentation for Atmospheric Rivers

Abstract The identification of atmospheric rivers (ARs) is crucial for weather and climate predictions as they are often associated with severe storm systems and extreme precipitation, which can cause large impacts on society. This study presents a deep learning model, termed ARDetect, for image segmentation of ARs using ERA5 data from 1960 to 2020 with labels obtained from the TempestExtremes tracking algorithm. ARDetect is a convolutional neural network (CNN)-based U-Net model, with its structure having been optimized using automatic hyperparameter tuning. Inputs to ARDetect were selected to be the integrated water vapor transport (IVT) and total column water (TCW) fields, as well as the AR mask from TempestExtremes from the previous time step to the one being considered. ARDetect achieved a mean intersection-over-union (mIoU) rate of 89.04% for ARs, indicating its high accuracy in identifying these weather patterns and a superior performance than most deep learning–based models for AR detection. In addition, ARDetect can be executed faster than the TempestExtremes method (seconds vs minutes) for the same period. This provides a significant benefit for online AR detection, especially for high-resolution global models. An ensemble of 10 models, each trained on the same dataset but having different starting weights, was used to further improve on the performance produced by ARDetect, thus demonstrating the importance of model diversity in improving performance. ARDetect provides an effective and fast deep learning–based model for researchers and weather forecasters to better detect and understand ARs, which have significant impacts on weather-related events such as floods and droughts.

Galea, Daniel↗

Machine Learning (ML) Classifier to Assist Metadata Creation

The Atmospheric Radiation Measurement (ARM) Data Center is responsible for the timely collection, archival, and curation of science data products. These products are freely available through an online data repository. Metadata creation is paramount for scientific users to find and access over seven petabytes of atmospheric science data. The hierarchical metadata structure allows users to search for information at both broad and narrow levels. This project aims to leverage 30 years’ worth of manually created metadata to enable machine predictions of broad-term classifications from narrow-term descriptions. These classification predictions would assist metadata coordinators with their term selections. This paper discusses the cleaning and preprocessing of the training data, the pipeline developed to determine the best model for this task, and the creation of an API metadata classifier for ARM measurement metadata. Our results show that the Linear Support Vector Classification (LinearSVC) algorithm, along with the Term Frequency – Inverse Document Frequency (TF-IDF) vectorizer, is well-suited for our multi-class classification task. Lengthier input training data led to better results, and artificial balancing was unnecessary for this particular use case. This predictive classifier enhances efficiency in metadata creation, as well as supports greater consistency and accuracy in metadata tagging.

Collier, Hannah [ORNL] (ORCID:0000000341284292)↗

Hybrid Imitation Learning for Real-Time Service Restoration in Resilient Distribution Systems

Self-healing capability is a critical factor for a resilient distribution system, which requires intelligent agents to automatically perform service restoration online, including network reconfiguration and reactive power dispatch. Here, the article proposes the imitation learning framework for training such an agent, where the agent will interact with an expert built based on the mixed-integer program to learn its optimal policy, and therefore significantly improve the training efficiency compared with exploration-dominant reinforcement learning (RL) methods. This significantly improved training efficiency makes the training problem under N-k scenarios tractable. A hybrid policy network is proposed to handle tie-line operations and reactive power dispatch simultaneously to further improve the restoration performance. The 33-bus and 119-bus systems with N-k disturbances are employed to conduct the training. The results indicate that the proposed method outperforms traditional RL algorithms such as the deep-Q network.

42 ENGINEERING↗

A general Bayesian algorithm for the autonomous alignment of beamlines

Autonomous methods to align beamlines can decrease the amount of time spent on diagnostics, and also uncover better global optima leading to better beam quality. The alignment of these beamlines is a high-dimensional expensive-to-sample optimization problem involving the simultaneous treatment of many optical elements with correlated and nonlinear dynamics. Bayesian optimization is a strategy of efficient global optimization that has proved successful in similar regimes in a wide variety of beamline alignment applications, though it has typically been implemented for particular beamlines and optimization tasks. In this paper, we present a basic formulation of Bayesian inference and Gaussian process models as they relate to multi-objective Bayesian optimization, as well as the practical challenges presented by beamline alignment. We show that the same general implementation of Bayesian optimization with special consideration for beamline alignment can quickly learn the dynamics of particular beamlines in an online fashion through hyperparameter fitting with no prior information. We present the implementation of a concise software framework for beamline alignment and test it on four different optimization problems for experiments on X-ray beamlines at the National Synchrotron Light Source II and the Advanced Light Source, and an electron beam at the Accelerator Test Facility, along with benchmarking on a simulated digital twin. We discuss new applications of the framework, and the potential for a unified approach to beamline alignment at synchrotron facilities.

47 OTHER INSTRUMENTATION↗

Future SDN-HPON Control Plane Architecture and Protocol for On-Demand Terabit End-to-End Extreme-Scale Science Applications

The UC Davis team conducted architecture, algorithm and experimental studies for next generation ultra-high-bandwidth optical networks in support of extreme-scale science applications. In particular, over the course of this 3-year project, the UC Davis team achieved the following main accomplishments: Designed a network architecture for automated multi-domain service provisioning: Designed information exchange schemes between autonomous systems (domains) Designed application programming interfaces to orchestrate resource reservation in data centers, high performance computation facilities, and communication networks belonging to different autonomous systems Experimentally assessed the multi-domain network architecture in a distributed field trial set-up connecting premises in three continents Designed a broker-based orchestrated multi-domain optical networking framework supporting observe-analyze-act cycle based cognitive inter-domain service provisioning: Designed an inter-domain operation mechanism, e.g., topology abstraction, for the broker plane and domain managers to collaboratively accomplish inter-domain service provisioning while respecting the autonomy of each administrative domain Designed an alien wavelength monitoring scheme and an artificial neural network based optical signal-to-noise ratio estimator for quality-of-transmission aware inter-domain service provisioning Developed a deep neural network based multi-domain traffic predictor and a cognitive routing and spectrum assignment algorithm to realize knowledge-based autonomous traffic engineering in multi-domain networks Implemented an optical signal-to-noise ratio estimator and traffic predictor with Google’s Tensor flow and assessed their performance using experimental and synthetic data Demonstrated the benefits of the proposed cognitive schemes with experiments ona two-domain elastic optical network testbed and evaluations on a six-domain simulation framework Built a two-domain software-defined networking control plane testbed using the ONOS platform and demonstrated the workflow of broker and domain manager cooperation, including hierarchical performance monitoring, topology abstraction and cognitive resource allocation. Designed self-learning cognitive service provisioning schemes using advanced machine learning technologies: Designed a self-taught anomaly detection framework for optical networks, including an unsupervised data clustering module for pattern recognition, a supervised data regression and classification module for online anomaly detection, and the system operation principle Experimentally validated the effectiveness of the self-taught anomaly detection framework with various failure/anomaly scenarios. Designed a self-learning autonomic routing, modulation and spectrum assignment (RMSA) provisioning framework for elastic optical networks with deep reinforcement learning. Designed effective modeling and training mechanisms for the self-learning RMSA framework Conducted extensive numerical simulations with various traffic patterns and topologies to verify the advantage of the self-learning RMSA framework.

42 ENGINEERING↗

Machine learning on FPGA for event selection

Real-time data processing is a frontier field in experimental particle physics. The application of FPGAs at the trigger level is used by many current and planned experiments (CMS, LHCb, Belle2, PANDA). Usually they use conventional processing algorithms. LHCb has implemented Machine Learning (ML) elements for real-time data processing with a triggered readout system that runs most of the ML algorithms on a computer farm. The work described in this article aims to test the ML-FPGA algorithms for streaming data acquisition. Herein, there are many experiments working in this area and they have a lot in common, but there are many specific solutions for detector and accelerator parameters that are worth exploring further. This report describes the purpose of the work and progress in evaluating the ML-FPGA application.

47 OTHER INSTRUMENTATION↗

RAP: Resource-aware Automated GPU Sharing for Multi-GPU Recommendation Model Training and Input Preprocessing

Ensuring high-quality recommendations for newly onboarded users requires the continuous retraining of Deep Learning Recommendation Models (DLRMs) with freshly generated data. To serve the online DLRM retraining, existing solutions use hundreds of CPU computing nodes designated for input preprocessing, causing significant power consumption that surpasses even the power usage of GPU trainers. To this end, we propose RAP, an end-to-end DLRM training framework that supports Resource-aware Automated GPU sharing for DLRM input Preprocessing and Training. The core idea of RAP is to accurately capture the remaining GPU computing resources during DLRM training for input preprocessing, achieving superior training efficiency without requiring additional resources. Specifically, RAP utilizes a co-running cost model to efficiently assess the costs of various input preprocessing operations, and it implements a resource-aware horizontal fusion technique that adaptively merges smaller kernels according to GPU availability, circumventing any interference with DLRM training. In addition, RAP leverages a heuristic searching algorithm that jointly optimizes both the input preprocessing graph mapping and the co-running schedule to maximize the end-to-end DLRM training throughput. The comprehensive evaluation shows that RAP achieves 78.3× speedup on average over CPU-based DLRM input preprocessing frameworks. In addition, the end-to-end training throughput of RAP is only 2.04% lower than the ideal case, which has no input preprocessing overhead.

Wang, Zheng↗