SEARCH · Engineering Papers
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Linux kernel driver for the Cray ClusterStor E1000 NVMe slot LEDs. This driver allows us to set the 'fault' and 'locate' LEDs on a CrayE1000 slot to indicate which drive has failed. Is there
Abstract not provided.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
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.
Training in Deep learning (DL) remains highly compute- and data-intensive, with I/O becoming a critical bottleneck as models and datasets scale. Recent studies report that data loading can dominate training time, especially on large-scale HPC systems with shared parallel file systems (PFS). Existing caching approaches either rely on single-tier designs or require intrusive modifications to training pipelines, limiting their portability and effectiveness. In this work, we present FitCache, a transparent drop-in framework for multi-tier caching to accelerate distributed DL training by coordinating fast local memory (e.g., DRAM, Persistent Memory (PMem)) and NVMe as hierarchical caches atop PFS. Our design adapts to hardware diversity, i.e., if NVMe is missing, memory transparently acts as a caching tier, ensuring stable performance. FitCache transparently intercepts I/O requests and issues concurrent fetches across all tiers, returning data from the fastest responder without centralized metadata or static redirection paths. FitCache adapts to dynamic workloads and heterogeneous clusters while maintaining POSIX compatibility. Experiments on Frontier (2048 GPUs) and smaller research clusters show that FitCache reduces training time by up to 40% and per-batch I/O latency by up to 71.6% compared to Lustre Orion PFS, offering a drop-in solution for scalable DL training.
In this paper, we propose OctoKV, an innovative network-based key-value storage system. OctoKV addresses the repetitive address translation overhead associated with traditional key-value stores running on file systems on the client side. To mitigate this overhead, we implemented the key-value store on the server side using NVMe-oF and a user-level NVMe driver. In particular, we employed fine-grained resource monitoring and load balancing based on heuristics to optimize I/O performance. OctoKV is deployed on a Linux cluster with Intel SPDK. The extensive evaluation shows that OctoKV achieves lower I/O response times in comparison to traditional approaches where key-value stores run on the client side. Also, the proposed load balancing strategies efficiently enhance I/O response times by equally distributing the workload from overloaded cores to other cores.
Fast 3D data analysis and steering of a tomographic experiment by changing environmental conditions or acquisition parameters require fast, close to real-time, 3D reconstruction of large data volumes. Here a performance-optimized TomocuPy package is presented as a GPU alternative to the commonly used central processing unit (CPU) based TomoPy package for tomographic reconstruction. TomocuPy utilizes modern hardware capabilities to organize a 3D asynchronous reconstruction involving parallel read/write operations with storage drives, CPU–GPU data transfers, and GPU computations. In the asynchronous reconstruction, all the operations are timely overlapped to almost fully hide all data management time. Since most cameras work with less than 16-bit digital output, the memory usage and processing speed are furthermore optimized by using 16-bit floating-point arithmetic. As a result, 3D reconstruction with TomocuPy became 20–30 times faster than its multi-threaded CPU equivalent. Full reconstruction (including read/write operations and methods initialization) of a 2048 3 tomographic volume takes less than 7 s on a single Nvidia Tesla A100 and PCIe 4.0 NVMe SSD, and scales almost linearly increasing the data size. To simplify operation at synchrotron beamlines, TomocuPy provides an easy-to-use command-line interface. Efficacy of the package was demonstrated during a tomographic experiment on gas-hydrate formation in porous samples, where a steering option was implemented as a lens-changing mechanism for zooming to regions of interest.
After using ROOT’s TTree I/O subsystem for over two decades and storing more than an exabyte of compressed High Energy Physics (HEP) data, advances in technology have motivated a complete redesign, RNTuple, which breaks backward-compatibility to take better advantage of these storage options. The RNTuple I/O subsystem has been designed to address performance bottlenecks and other shortcomings of TTree. Specifically, RNTuple comes with an updated, more compact binary data format that can be stored both in ROOT files and natively in object stores. It is designed for modern storage hardware (e.g. high-throughput low-latency NVMe SSDs), and provides robust and easy to use interfaces. The binary format of RNTuple is scheduled to become production grade in 2024, and recently has become mature enough to start exploring the integration into software used by HEP experiments. In this contribution, we discuss the developments to support the features as required by the ATLAS analysis Event Data Model (EDM) in RNTuple, which will enable its integration into the Athena software framework. With these developments in place, we evaluate the performance of the current most recent versions of RNTuple-based ATLAS data sets and compare this to that of TTree.
The RNTuple I/O subsystem is ROOT’s future event data file format and access API. It is driven by the expected data volume increase at upcoming HEP experiments, e.g. at the HL-LHC, and recent opportunities in the storage hardware and software landscape such as NVMe drives and distributed object stores. RNTuple is a redesign of the TTree binary format and API and has shown to deliver substantially faster data throughput and better data compression both compared to TTree and to industry standard formats. In order to let HENP computing workflows benefit from RNTuple’s superior performance, however, the I/O stack needs to connect efficiently to the rest of the ecosystem, from grid storage to (distributed) analysis frameworks to (multithreaded) experiment frameworks for reconstruction and ntuple derivation. With the RNTuple binary format soon arriving at its first production release, we present RNTuple’s feature set, integration efforts, and its performance impact on the time-to-solution. We show the latest performance figures of RDataFrame analysis code of realistic complexity, comparing RNTuple and TTree as data sources. We discuss RNTuple’s approach to functionality critical to the HENP I/O (such as multithreaded writes, fast data merging, schema evolution) and we provide an outlook on the road to its use in production.
For several years, the ROOT team is developing the new RNTuple I/O subsystem in preparation of the next generation of collider experiments. Both HL-LHC and DUNE are expected to start data taking by the end of this decade. They pose unprecedented challenges to event data I/O in terms of data rates, event sizes, and event complexity. At the same time, the I/O landscape is becoming more diverse. HPC cluster file systems and object stores, NVMe disk cache layers in analysis facilities, and S3 storage on cloud resources are mixing with traditional XRootD-managed spinning disk pools.The ROOT team will finalize a first production version of the RNTuple binary format by the end of 2024. After this point, ROOT will provide backward compatibility for RNTuple data. This contribution provides an overview of the RNTuple feature set, the related R&D activities and the long-term vision for RNTuple. We report on performance, interface design, tooling, robustness, integration with experiment frameworks, and validation results, as well as recent R&D on parallel reading and writing and exploitation of modern hardware and storage systems. We will give an outlook on possible future features after a first production release.Collaboratively, the IT and EP departments at CERN have launched a formal project within the Research and Computing sector to evaluate the novel data format for physics analysis data utilized in LHC experiments and other fields. This part of the project focuses on validating the scalability of the EOS storage backend during the transition from the over 25 years old TTree production format to the newly developed RNTuple format, using both replicated and erasure-coded storage profiles.
GNU Parallel is a versatile and powerful tool for process parallelization widely used in scientific computing. This paper demonstrates its effective application in high-performance computing (HPC) environments, particularly focusing on its scalability and efficiency in executing large-scale high-throughput high-performance computing (HT-HPC) workflows. Through real-world examples, we highlight GNU Parallel’s performance across various HPC workloads, including GPU computing, container-based workloads, and node-local NVMe storage. Our results on two leading supercomputers, OLCF’s Frontier and NERSC’s Perlmutter, showcase GNU Parallel’s rapid process dispatching ability and its capacity to maintain low overhead even at extreme scales. We explore GNU Parallel’s application in massive parallel file transfers using a scheduled Data Transfer Node (DTN) cluster, emphasizing its broad utility in diverse scientific workflows. Beyond its direct application as a viable workflow manager, GNU Parallel can be employed in conjunction with other workflow systems as a "last-mile" parallelizing driver and as a quick prototyping tool to design and extract parallel profiles from application executions. We then argue that the potential for GNU Parallel to transform workflow management at extreme scales is substantial, paving the way for more efficient and effective scientific discoveries.
Graph neural networks (GNNs) are a class of Deep Learning models used in designing atomistic materials for effective screening of large chemical spaces. To ensure robust prediction, GNN models must be trained on large volumes of atomistic data on leadership class supercomputers. Even with the advent of modern architectures that consist of multiple storage layers that include node-local NVMe devices in addition to device memory for caching large datasets, extreme-scale model training faces I/O challenges at scale.We present DDStore, an in-memory distributed data store designed for GNN training on large-scale graph data. DDStore provides a hierarchical, distributed, data caching technique that combines data chunking, replication, low-latency random access, and high throughput communication. DDStore achieves near-linear scaling for training a GNN model using up to 1000 GPUs on the Summit and Perlmutter supercomputers, and reaches up to a 6.15x reduction in GNN training time compared to state-of-the-art methodologies.
As the name suggests, The Operando Neutron Measurement of Metal Additive Manufacturing project conducted by ORNL’s Manufacturing Demonstration Facility (MDF) is experimenting with advanced additive metal manufacturing techniques while analyzing the process using the Spallation Neutron Source’s (SNS) beamline. As part of this experiment, the MDF is seeking to employ 2 XIMEA visible light cameras and a single infrared camera to analyze and correct manufacturing in real-time. The MDF requires a solution for capturing the high-resolution data feed from the cameras with compression while preserving enough detail for their software to detect and correct errors in real time. Our solution was to develop a Robot Operating System (ROS) driver to feed the camera data into ROS. From ROS, the feed is compressed and temporarily stored locally to a stripped 4 NVMe SSD RAID array. Post-experiment, the data is transitioned to long-term storage for archival purposes.
Automotive and industrial grade SSDs were tested for TID and SEE response at the assembly level to investigate radiation tolerance trends and explore radiation hardness assurance best practices in commercial memory devices. SSDs were installed in passive NVMe extenders to place only the drive in the beam line. A digital I/O module connected to the test computer provided inhibit signals to block both facility beam delivery and power while attempting recovery from any device failure conditions (e.g., failed write, failed read, or unresponsive device).
The driving goal is to investigate the feasibility of a message-based data transport for SPDK. All of SPDK’s existing data transport protocols are “connection-based”, requiring a target to hold state for each connected client / initiator. At extreme scale, juggling these connections from a large number of initiators creates a performance bottleneck. Under a message-based paradigm, we can bypass this behavior, as all initiators share a global communication channel, held open for the entire run. In theory, this approach should minimize connection management overhead at extreme scale.