Engineering topics

Duplyakin, Dmitry

Publications and source records attributed to Duplyakin, Dmitry.

At least 19 records

DOE OSTI2024

Generic and ML Workloads in an HPC Datacenter: Node Energy, Job Failures, and Node-Job Analysis

HPC datacenters offer a backbone to the modern digital society. Increasingly, they run Machine Learning (ML) jobs next to generic, compute-intensive workloads, supporting science, business, and other decision-making processes. However, understanding how ML jobs impact the operation of HPC datacenters, relative to generic jobs, remains desirable but understudied. In this work, we leverage long-term operational data, collected from a national-scale production HPC datacenter, and statistically compare how ML and generic jobs can impact the performance, failures, resource utilization, and energy consumption of HPC datacenters. Our study provides key insights, e.g., ML-related power usage causes GPU nodes to run into temperature limitations, median/mean runtime and failure rates are higher for ML jobs than for generic jobs, both ML and generic jobs exhibit highly variable arrival processes and resource demands, significant amounts of energy are spent on unsuccessfully terminating jobs, and concurrent jobs tend to terminate in the same state. We open-source our cleaned-up data traces on Zenodo (https://doi. org/10.5281/zenodo.13685426), and provide our analysis toolkit as software hosted on GitHub (https://github.com/atlarge-research/2024-icpads-hpc-workload-characterization). This study offers multiple benefits for data center administrators, who can improve operational efficiency, and for researchers, who can further improve system designs, scheduling techniques, etc.

crossanalysis↗

DOE OSTI2024

Tandem Predictions for HPC Jobs

At the core of the predictive analytics applied to High Performance Computing (HPC), the most prominent tasks are the prediction of job runtimes and the prediction of job queue times, both of which have the potential for informing HPC users during their every-day decision making. Accurate runtime predictions can help users better choose so-called wallclock times at job submission, decreasing the odds of their jobs waiting in queues longer than necessary. The accurate and timely queue time predictions offered for the available partitions can inform the favorable selection of partitions for running jobs. This potential is well understood as we see in the abundance of research studies that propose solutions for these tasks, including the work published in the last several years. These tasks are seemingly receptive to the Machine Learning (ML) solutions, considering that there is no shortage of training data where HPC centers over time run millions and millions of jobs. However, we study the existing research literature, as well as look for examples in the toolchains supported on the exemplar HPC facilities, and, surprisingly, do not find any practical solutions that are ready to be adopted. We interpret this as a manifestation of the shortage of UX/UI efforts that support HPC analytics and also as a sign that the research has not come to the consensus on solving these tasks. In this study, we aim to shed new light on the long-running task of job queue time prediction by exploring the utility of runtime predictions in improving prediction accuracy and, actually, predicting these two metrics together, in tandem. In other words, we show how runtime predictions become valuable input in the queue time modeling. We challenge the existing approaches to feature engineering for the queue time prediction and describe promising results we obtained for a large dataset of HPC jobs from a supercomputer at the National Renewable Energy Laboratory.

HPC↗

DOE OSTI2024

Quantifying Uncertainty in HPC Job Queue Time Predictions

High Performance Computing (HPC) has developed at an unprecedented pace in recent decades. This growth has demanded corresponding development in the area of HPC Operational Data Analytics (ODA), which encompasses a wide range of data analysis techniques, ML/AI efforts, tools, and visualizations. Published studies in ODA offer a variety of practical ways to inform HPC users, administrators, procurement managers, and other stakeholders. Uncertainty analysis, however, is rare in the related published literature. For instance, we identify only 1 out of 14 existing studies focused on job queue time prediction that investigates the uncertainty aspect of their proposed predictions. We recognize the utmost importance uncertainty quantification can have in such predictive analytics solutions, with consequences in how users interpret information they receive, and attempt to bridge this gap. With the goal of improving access to such insights, we develop a process for determining upper and lower bounds of the predicted queue times of a regression model at a specified confidence level. Our current research is focused on the uncertainty in predicting job queue times, yet our approach may be employed in predicting other metrics.

HPC↗

DOE OSTI2024

Geospatial Data Platform for All

Spatiotemporal data has evolved in scale due to augmented use in cross-domain applications. Simultaneously, there is substantial growth in the availability of Geographic Information Systems (GIS) data provided by the United States Geological Survey (USGS) along with other federal, state, county, or local agencies through open-data portals and public access APIs. However, data availability does not equate with accessibility. Large-scale analyses and applications require robust, performant data management with co-location of data storage and computing. The insufficiency of data management infrastructure compels researchers to adopt ad hoc project- specific GIS data storage solutions (e.g., copying data to High-Performance computer file systems). As an ad hoc storage strategy does not scale, it hampers cross-domain analyses causing difficulty in data reuse and utilizing existing code bases. Furthermore, GIS data is complex and requires expertise to analyze and manipulate due to its intricate data structures and data-specific projection transformations. Despite the challenges, we recognize that derived GIS data products, e.g., satellite or LIDAR-based images, can be used in downstream applications such as AI by domain, but non-GIS experts. To address the data needs and overcome the challenges, we are working towards a GIS Data Platform focused on efficient data storage, data discovery and access, and an API to enable common workflows. We propose a knowledge-graph (KG) approach for data discovery, whereby datasets are semantically linked to higher- level constructs such as projects and research areas. The semantic data links enable researchers to explore datasets in a top-down approach by specifying relevant and meaningful terms (assists in finding hidden data). An advantage is that the nodes and edges in a knowledge graph create built-in semantic documentation. Deeper spatiotemporal connections between data sources can be encoded via Graph Neural Networks (GNN) (Zhang et al., 2021). The KG approach can be extended to integrate the data itself in a Virtual KG (VKG). Our work will derive inspiration from large-scale VKG efforts that have been undertaken or are currently underway as part of the OpenStreetMap project (Ding et al., 2021). For DOE Data Days, we share the proposed geospatial data platform hybrid (cloud/on-prem) architecture, our work-to-date on storing, retrieving, and transforming LiDAR and raster data relevant to two important NREL use-cases, including the Renewable Energy Potential (reV) Model, and present our proposal for a KG based data discovery engine.

data platform↗

DOE OSTI2024

Resource Assessment for Distributed Wind Energy: An Evaluation of Best-Practice Methods in the Continental US

Current wind resources within the United States (US) indicate a potential to profitably install nearly 1,400 gigawatts of distributed wind (DW) capacity. This amount is equivalent to over half of the United States’ current energy demand from electricity, making it enough to power millions of homes and businesses and replace countless fossil fuel-based generating plants. Despite the potential growth of DW in the US, deployments are presently hindered by a lack of confidence in resource estimation methods. One potential challenge is that smaller-scale turbines, with hub heights of 40 meters or less, are disproportionately impacted by obstacles such as buildings and vegetation. These obstacles may produce complex wake effects, best modeled with high-fidelity complex fluid dynamics (CFD) models that are too computationally expensive to use for routine siting and resource assessment. Thus, installers today make use of heuristics and simple equations to approximate the impact of obstacles while also leveraging long-term resource data from commercial or publicly available atmospheric models. This study evaluates these historical and commonly used methods alongside new lower-order obstacle models produced from CFD simulations and measurement-based bias correction. The preliminary results from this study show the importance of taking care in the choice and application of mesoscale atmospheric models and the significant value of bias correction using measurements from nearby meteorological towers. Detailed obstacle modeling provides only modest additional gains in performance and, in some cases, can add error, especially at sites where turbines have already been located to avoid obvious impact from upwind obstacles. These findings reinforce the importance of collecting in situ measurements and suggest that obstacle models may be better applied in practice to automated or computer-aided siting, rather than in economic wind resource assessments.

17 WIND ENERGY↗

DOE OSTI2024

Tools Assessing Performance (TAP) 2.0

Dmitry Duplyakin will be presenting on the latest research and results in the Tools Assessing Performance (TAP) 2.0 project. This presentation will include updates on the latest data the group has produced, integration of obstacle models in the computational pipeline for distributed wind siting, and the plans for the near-term analysis and validation efforts. The talk will acknowledge the work of collaborators from NREL and three other national labs - ANL, LANL, and PNNL - all contributing to this multi-year project.

distributed wind↗

DOE OSTI2023

One System, Many Models: Designing a Surrogate Model for Sulfur Thermal Energy Storage: Preprint

Industrial process heating (IPH) relies primarily on thermal energy generated by fossil fuel combustion to produce, treat, and alter manufactured goods. Thermal energy storage (TES) helps reduce the carbon footprint of IPH systems by facilitating the utilization of renewable and waste heat sources. A promising new TES technology uses elemental sulfur as the heat-storage medium. The design of sulfur TES systems can be evaluated with the aid of computational fluid dynamics (CFD). However, the computational cost of such CFD efforts is prohibitive to comprehensive optimizations over design parameters. To reduce this computational cost, machine learning (ML) models can be developed to act as surrogates for CFD. In this paper, we describe the process of building and evaluating surrogate ML models for facilitating optimization of sulfur TES systems for IPH. To enforce the thermodynamic relationship between the two modeled quantities, we develop a hybrid model for sulfur temperature using both direct predictions of temperature and calculations of temperature from predictions of the heat transfer coefficient. The hybrid model enforces this constraint at the expense of the slightly reduced accuracy compared to two disjoint models. The overall high accuracy observed in our model evaluation demonstrates the usefulness of such surrogate modeling for studying TES systems. This work contributes to the field of TES surrogate modeling by offering a novel accurate hybrid approach to predicting simultaneously the heat transfer coefficient and temperature of the heat storage medium.

computational fluid dynamics↗

DOE OSTI2023

Mastering HPC Runtime Prediction: From Observing Patterns to a Methodological Approach

The continual expansion of high-performance computing (HPC) brings with it an increasing need for efficiency. Heavy investment in energy, hardware, and software infrastructure to support peta- and exascale computing requires the optimization of existing systems and, wherever possible, the discernment and adoption of best-practices towards these goals. Such is the case for runtime prediction. When a job is submitted to an HPC system, an estimate of its runtime is provided by the user in the form of "requested wallclock". Error in this user-provided estimate can lead to jobs being prematurely killed by the scheduler, increased wait time on the queue, and decreased system utilization. More than fifteen years of research has been directed at mitigating these effects by using data-driven runtime predictions. Codified here is a set of commonalities and insights emerging from this body of work, which we present as recommendations and best practices. These practices are combined into a methodological approach described and evaluated on an 11-million-job dataset from the National Renewable Energy Laboratory's petascale HPC system, Eagle. This dataset and the accompanying codebase have been released to the public domain for the benefit of the wider HPC research community.

high performance computing↗

DOE OSTI2023

Is Knowledge about Running Applications Helping Improve Runtime Prediction of HPC Jobs?

High-performance computing systems rely upon scheduling algorithms to achieve high utilization. These schedulers rely upon user estimates of job resource requirements, such as runtime, to determine optimal scheduling of incoming jobs. These user estimates, however, are prone to error. To mitigate this error, significant research has been directed at providing better estimates of job runtime, usually employing machine learning techniques. These techniques are dependent upon the input features selected. Among the possible features is the primary application used by the job. In a survey of more than 20 papers directed at improving runtime prediction, only four included primary application as an input feature. We focus this investigation specifically on the value of adding primary application as an input feature, and find that it does improve model performance, especially for jobs with longer runtimes, though this improvement varies based on the application used. We recommend further research to determine the cause of this variability as well as an optimal strategy for employing a mixture of models both including and not including primary application as a feature.

MATHEMATICS AND COMPUTING↗

DOE OSTI2023

Mastering HPC Runtime Prediction: From Observing Patterns to a Methodological Approach: Preprint

The continual expansion of high-performance computing (HPC) brings with it an increasing need for efficiency. Heavy investment in energy, hardware, and software infrastructure to support peta- and exascale computing requires the optimization of existing systems and, wherever possible, the discernment and adoption of best-practices towards these goals. Such is the case for runtime prediction. When a job is submitted to an HPC system, an estimate of its runtime is provided by the user in the form of "requested wallclock''. Error in this user-provided estimate can lead to jobs being prematurely killed by the scheduler, increased wait time on the queue, and decreased system utilization. More than fifteen years of research has been directed at mitigating these effects by using data-driven runtime predictions. Codified here is a set of commonalities and insights emerging from this body of work, which we present as recommendations and best practices. These practices are combined into a methodological approach described and evaluated on an 11-million-job dataset from the National Renewable Energy Laboratory's petascale HPC system, Eagle. This dataset and the accompanying codebase have been released to the public domain for the benefit of the wider HPC research community.

high performance computing↗

DOE OSTI2023

Is Knowledge About Running Applications Helping Improve Runtime Prediction of HPC Jobs?

feature selection↗

DOE OSTI2022

Sulfur TES ML [SWR-23-02]

Sulfur TES ML contains code originally created to develop surrogate ML models to mimic the utility of CFD simulations of Sulfur TES systems. It has been adapted to provide a comprehensive package for building, training, validating, testing, and optimizing regression models with customizable features and targets.

Menear, Kevin↗

DOE OSTI2022

End-to-end optimization for battery materials and molecules by combining graph neural networks and reinforcement learning

The National Renewable Energy Laboratory (NREL), together with the Colorado School of Mines (CSM) and Colorado State University (CSU), has developed a machine learning-enhanced approach to design new battery materials. Currently, such materials are designed in part via numerous expensive high-fidelity computational simulations that predict the performance of a given composition. Even with computational screening tools, the vast landscape of possible molecular or crystal structures exceeds current and future computational capacity. Improving the efficiency by which new materials can be optimized will therefore disrupt the cost, risk, and time required to bring new energy solutions to the marketplace. Predicting the properties of an organic molecule or periodic crystalline material given its structure has grown increasingly common. These approaches leverage large-scale computational and experimental databases and ML approaches such as graph neural networks. The inverse design problem of finding a material that possesses desired properties is substantially more challenging, since enumerating all valid material structures is not feasible. In this project, we leveraged recent success in reinforcement learning to efficiently navigate this high-dimensional search space. Just as algorithms can find the optimal chess moves from nearly limitless options, we train an approach to evolve a simple starting structure into a complex structure that possess the desired properties. Our solution has been demonstrated by applying it to two related design application tasks for short- and long-term energy storage, respectively: (1) the design of solid-state ion conductors and (2) the design of organic redox-active materials. The project has resulted an open-source software library for material design, documented examples of applying the library to both organic and inorganic material optimization, and peer-reviewed publications detailing the data, computational models, and resulting candidate materials.

25 ENERGY STORAGE↗

DOE OSTI2022

End-to-End Optimization for Battery Materials and Molecules by Combining Graph Neural Networks and Reinforcement Learning

The National Renewable Energy Laboratory (NREL), together with the Colorado School of Mines (CSM) and Colorado State University (CSU), has developed a machine learning-enhanced approach to the design of new battery materials. Currently, such materials are designed in part via numerous expensive high-fidelity computational simulations that predict the performance of a given composition. Even with computational screening tools, the vast landscape of possible molecular or crystal structures exceeds current and future computational capacity. Improving the efficiency by which new materials can be optimized will therefore disrupt the cost, risk, and time required to bring new energy solutions to the marketplace. Predicting the properties of an organic molecule or periodic crystalline material given its structure has grown increasingly common. These approaches leverage large-scale computational and experimental databases and ML approaches such as graph neural networks. The inverse design problem of finding a material that possesses desired properties is substantially more challenging, since enumerating all valid material structures is not feasible. In this project, we leveraged recent success in reinforcement learning to efficiently navigate this high-dimensional search space. Just as algorithms can find the optimal chess moves from nearly limitless options, we train an approach to evolve a simple starting structure into a complex structure that possess the desired properties. Our solution has been demonstrated by applying it to two related design application tasks for short- and long-term energy storage, respectively: (1) the design of solid-state ion conductors and (2) the design of organic redox-active materials. The project has resulted an open-source software library for material design, documented examples of applying the library to both organic and inorganic material optimization, and peer-reviewed publications detailing the data, computational models, and resulting candidate materials.

25 ENERGY STORAGE↗

DOE OSTI2022

graphenv: a Python library for reinforcement learning on graph search spaces

Many important and challenging problems in combinatorial optimization (CO) can be expressed as graph search problems, in which graph vertices represent full or partial solutions and edges represent decisions that connect them. Graph structure not only introduces strong relational inductive biases for learning (Battaglia et al., 2018) - in this context, by providing a way to explicitly model the value of transitioning (along edges) between one search state (vertex) and the next - but lends itself to problems both with and without clearly defined algebraic structure. For example, classic CO problems on graphs such as the Traveling Salesman Problem (TSP) can be expressed as either pure graph search or integer programs. Other problems, however, such as molecular optimization, do no have concise algebraic formulations and yet are readily implemented as a graph search (V. et al., 2022; Zhou et al., 2019). Such "model-free" problems constitute a large fraction of modern reinforcement learning (RL) research owing to the fact that it is often much easier to write a forward simulation that expresses all of the state transitions and rewards, than to write down the precise mathematical expression of the full optimization problem. In the case of molecular optimization, for example, one can use domain knowledge alongside existing software libraries to model the effect of adding a single bond or atom to an existing but incomplete molecule, and let the RL algorithm build a model of how good a given decision is by "experiencing" the simulated environment many times through. In contrast, a model-based mathematical formulation that fully expresses all the chemical and physical constraints is intractable. In recent years, RL has emerged as an effective paradigm for optimizing searches over graphs and led to state-of-the-art heuristics for games like Go and chess, as well as for classical CO problems such as the TSP. This combination of graph search and RL, while powerful, requires non-trivial software to execute, especially when combining advanced state representations such as Graph Neural Networks (GNN) with scalable RL algorithms.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2022

Multi-objective goal-directed optimization of de novo stable organic radicals for aqueous redox flow batteries

Abstract Advances in the field of goal-directed molecular optimization offer the promise of finding feasible candidates for even the most challenging molecular design applications. One example of a fundamental design challenge is the search for novel stable radical scaffolds for an aqueous redox flow battery that simultaneously satisfy redox requirements at the anode and cathode, as relatively few stable organic radicals are known to exist. To meet this challenge, we develop a new open-source molecular optimization framework based on AlphaZero coupled with a fast, machine-learning-derived surrogate objective trained with nearly 100,000 quantum chemistry simulations. The objective function comprises two graph neural networks: one that predicts adiabatic oxidation and reduction potentials and a second that predicts electron density and local three-dimensional environment, previously shown to be correlated with radical persistence and stability. With no hard-coded knowledge of organic chemistry, the reinforcement learning agent finds molecule candidates that satisfy a precise combination of redox, stability and synthesizability requirements defined at the quantum chemistry level, many of which have reasonable predicted retrosynthetic pathways. The optimized molecules show that alternative stable radical scaffolds may offer a unique profile of stability and redox potentials to enable low-cost symmetric aqueous redox flow batteries.

25 ENERGY STORAGE↗

DOE OSTI2022

Evaluation of obstacle modelling approaches for resource assessment and small wind turbine siting: case study in the northern Netherlands

Abstract. Growth in adoption of distributed wind turbines for energy generation is significantly impacted by challenges associated with siting and accurate estimation of the wind resource. Small turbines, at hub heights of 40 m or less, are greatly impacted by terrestrial obstacles such as built structures and vegetation that can cause complex wake effects. While some progress in high-fidelity complex fluid dynamics (CFD) models has increased the potential accuracy for modelling the impacts of obstacles on turbulent wind flow, these models are too computationally expensive for practical siting and resource assessment applications. To understand the efficacy of available models in situ, this study evaluates classic and commonly used methods alongside new state-of-the-art lower-order models derived from CFD simulations and machine learning approaches. This evaluation is conducted using a subset of an extensive original dataset of measurements from more than 300 operational wind turbines in the northern Netherlands. The results show that data-driven methods (e.g. machine learning and statistical modelling) are most effective at predicting production at real sites with an average error in annual energy production of 2.5 %. When sufficient data may not be available de novo to support these data-driven approaches, models derived from high-fidelity simulations show promise and reliably outperform classic methods. On average these models have 6.3 %–11.5 % error compared with 26 % for classic methods and 27 % baseline error for reanalysis data without obstacle correction. While more performant on average, these methods are also sensitive to the quality of obstacle descriptions and reanalysis inputs.

17 WIND ENERGY↗

DOE OSTI2022

Graph-Env

The Graph-Env library provides a framework for adapting graph search problems into OpenAI Gym environments for reinforcement learning. In other words, Graph-Env enables reinforcement learning algorithms to be applied to graph search problems. Graph search problems include molecule and crystal structure design problems; route planning problems, including traveling salesperson problem; puzzles and games, such as chess and go; shortest path problems; minimum spanning tree; vehicle path generation; and others.

Tripp, Charles↗