Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Parallel I/O”

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 91 records · Page 5
All sourcesNASA NTRSDOE OSTI
NASA NTRS1996

The Galley Parallel File System

Most current multiprocessor file systems are designed to use multiple disks in parallel, using the high aggregate bandwidth to meet the growing I/0 requirements of parallel scientific applications. Many multiprocessor file systems provide applications with a conventional Unix-like interface, allowing the application to access multiple disks transparently. This interface conceals the parallelism within the file system, increasing the ease of programmability, but making it difficult or impossible for sophisticated programmers and libraries to use knowledge about their I/O needs to exploit that parallelism. In addition to providing an insufficient interface, most current multiprocessor file systems are optimized for a different workload than they are being asked to support. We introduce Galley, a new parallel file system that is intended to efficiently support realistic scientific multiprocessor workloads. We discuss Galley's file structure and application interface, as well as the performance advantages offered by that interface.

Nieuwejaar, Nils↗
DOE OSTI2022

Large Scale Caching and Streaming of Training Data for Online Deep Learning

The training of deep neural network models on large data remains a difficult problem, despite progress towards scalable techniques. In particular, there is a mismatch between the random but predetermined order in which AI flows select training samples and the streaming I/O patterns for which traditional HPC data storage (e.g., parallel file systems) are designed. In addition, as more data are obtained, it is feasible neither simply to train learning models incrementally, due to catastrophic forgetting (i.e., bias towards new samples), nor to train frequently from scratch, due to prohibitive time and/or resource constraints. In this paper, we study data management techniques that combine caching and streaming with rehearsal support in order to enable efficient access to training samples in both offline training and continual learning. We revisit state-of-art streaming approaches based on data pipelines that transparently handle prefetching, caching, shuffling, and data augmentation, and discuss the challenges and opportunities that arise when combining these methods with data-parallel training techniques. We also report on preliminary experiments that evaluate the I/O overheads involved in accessing the training samples from a parallel file system (PFS) under several concurrency scenarios, highlighting the impact of the PFS on the design of the data pipelines.

data pipelines↗
DOE OSTI2021

I/O Bottleneck Detection and Tuning: Connecting the Dots using Interactive Log Analysis

Using parallel file systems efficiently is a tricky problem due to inter-dependencies among multiple layers of I/O software, including high-level I/O libraries (HDF5, netCDF, etc.), MPI-IO, POSIX, and file systems (GPFS, Lustre, etc.). Profiling tools such as Darshan collect traces to help understand the I/O performance behavior. However, there are significant gaps in analyzing the collected traces and then applying tuning options offered by various layers of I/O software. Seeking to connect the dots between I/O bottleneck detection and tuning, we propose DXT Explorer, an interactive log analysis tool. In this paper, we present a case study using our interactive log analysis tool to identify and apply various I/O optimizations. We report an evaluation of performance improvement achieved for four I/O kernels extracted from science applications.

Bez, Jean Luca↗
DOE OSTI2022

Machine Learning Assisted HPC Workload Trace Generation for Leadership Scale Storage Systems

Monitoring and analyzing a wide range of I/O activities in an HPC cluster is important in maintaining mission-critical performance in a large-scale, multi-user, parallel storage system. Center-wide I/O traces can provide high-level information and fine-grained activities per application or per user running in the system. Studying such large-scale traces can provide helpful insights into the system. It can be used to develop predictive methods for making predictive decisions, adjusting scheduling policies, or providing decisions for the design of next-generation systems. However, sharing real-world I/O traces to expedite such research efforts leaves a few concerns; i) the cost of sharing the large traces is expensive due to this large size, and ii) privacy concern is an issue.We address such issues by building an end-to-end machine learn- ing (ML) workflow that can generate I/O traces for large-scale HPC applications. We leverage ML based feature selection and gener- ative models for I/O trace generation. The generative models are trained on I/O traces collected by the darshan I/O characterization tool over a period of one year. We present a two-step generation process consisting of two deep-learning models, called the feature generator and the trace generator. The combination of two-step generative models provides robustness by reducing the bias of the model and accounting for the stochastic nature of the I/O traces across different runs of an application. We evaluate the performance of the generative models and show that the two-step model can generate time-series I/O traces with less than 20% root mean square error.

Paul, Arnab↗
NASA NTRS1990

A highly reliable, autonomous data communication subsystem for an advanced information processing system

The need to meet the stringent performance and reliability requirements of advanced avionics systems has frequently led to implementations which are tailored to a specific application and are therefore difficult to modify or extend. Furthermore, many integrated flight critical systems are input/output intensive. By using a design methodology which customizes the input/output mechanism for each new application, the cost of implementing new systems becomes prohibitively expensive. One solution to this dilemma is to design computer systems and input/output subsystems which are general purpose, but which can be easily configured to support the needs of a specific application. The Advanced Information Processing System (AIPS), currently under development has these characteristics. The design and implementation of the prototype I/O communication system for AIPS is described. AIPS addresses reliability issues related to data communications by the use of reconfigurable I/O networks. When a fault or damage event occurs, communication is restored to functioning parts of the network and the failed or damage components are isolated. Performance issues are addressed by using a parallelized computer architecture which decouples Input/Output (I/O) redundancy management and I/O processing from the computational stream of an application. The autonomous nature of the system derives from the highly automated and independent manner in which I/O transactions are conducted for the application as well as from the fact that the hardware redundancy management is entirely transparent to the application.

Nagle, Gail↗
DOE OSTI2021

Characterizing Machine Learning I/O Workloads on Leadership Scale HPC Systems

High performance computing (HPC) is no longer solely limited to traditional workloads such as simulation and modeling. With the increase in the popularity of machine learning (ML) and deep learning (DL) technologies, we are observing that an increasing number of HPC users are incorporating ML methods into their workflow and scientific discovery processes, across a wide spectrum of science domains such as biology, earth science, and physics. This gives rise to a diverse set of I/O patterns than the traditional checkpoint/restart-based HPC I/O behavior. The details of the I/O characteristics of such ML I/O workloads have not been studied extensively for large-scale leadership HPC systems. This paper aims to fill that gap by providing an in-depth analysis to gain an understanding of the I/O behavior of ML I/O workloads using darshan - an I/O characterization tool designed for lightweight tracing and profiling. We study the darshan logs of more than 23, 000 HPC ML I/O jobs over a time period of one year running on Summit - the second-fastest supercomputer in the world. This paper provides a systematic I/O characterization of ML I/O jobs running on a leadership scale supercomputer to understand how the I/O behavior differs across science domains and the scale of workloads, and analyze the usage of parallel file system and burst buffer by ML I/O workloads.

Paul, Arnab↗
NASA NTRS1995

Exploring the use of I/O nodes for computation in a MIMD multiprocessor

As parallel systems move into the production scientific-computing world, the emphasis will be on cost-effective solutions that provide high throughput for a mix of applications. Cost effective solutions demand that a system make effective use of all of its resources. Many MIMD multiprocessors today, however, distinguish between 'compute' and 'I/O' nodes, the latter having attached disks and being dedicated to running the file-system server. This static division of responsibilities simplifies system management but does not necessarily lead to the best performance in workloads that need a different balance of computation and I/O. Of course, computational processes sharing a node with a file-system service may receive less CPU time, network bandwidth, and memory bandwidth than they would on a computation-only node. In this paper we begin to examine this issue experimentally. We found that high performance I/O does not necessarily require substantial CPU time, leaving plenty of time for application computation. There were some complex file-system requests, however, which left little CPU time available to the application. (The impact on network and memory bandwidth still needs to be determined.) For applications (or users) that cannot tolerate an occasional interruption, we recommend that they continue to use only compute nodes. For tolerant applications needing more cycles than those provided by the compute nodes, we recommend that they take full advantage of both compute and I/O nodes for computation, and that operating systems should make this possible.

Kotz, David↗
NASA NTRS1986

Application of data flow concepts to a multigrid solver for the Euler equations

In this study a multigrid solver for Euler equations (FLO52R) was examined to determine its performance potential on a hypothetical computer using a data flow architecture. The proposed computer would require massive parallelism to realize its design performance. On the other hand this parallelism would be more easily realized than with a conventional vector processor such as the Cray-1S. Several changes to the proposed design substantially alleviated most of the remaining bottlenecks to parallel processing. Other changes allowed clearer definition of memory access and disk I/O. Finally, a portion of the algorithm was rewritten to improve parallel performance. With these changes, performance levels approaching that of a Cray-1S may be possible for a computer costing far less. Estimates are given for overall speed, memory, and network bandwidth, and for instruction memory requirements.

Merriam, M. L.↗
NASA NTRS1995

High Performance Parallel Computational Nanotechnology

At a recent press conference, NASA Administrator Dan Goldin encouraged NASA Ames Research Center to take a lead role in promoting research and development of advanced, high-performance computer technology, including nanotechnology. Manufacturers of leading-edge microprocessors currently perform large-scale simulations in the design and verification of semiconductor devices and microprocessors. Recently, the need for this intensive simulation and modeling analysis has greatly increased, due in part to the ever-increasing complexity of these devices, as well as the lessons of experiences such as the Pentium fiasco. Simulation, modeling, testing, and validation will be even more important for designing molecular computers because of the complex specification of millions of atoms, thousands of assembly steps, as well as the simulation and modeling needed to ensure reliable, robust and efficient fabrication of the molecular devices. The software for this capacity does not exist today, but it can be extrapolated from the software currently used in molecular modeling for other applications: semi-empirical methods, ab initio methods, self-consistent field methods, Hartree-Fock methods, molecular mechanics; and simulation methods for diamondoid structures. In as much as it seems clear that the application of such methods in nanotechnology will require powerful, highly powerful systems, this talk will discuss techniques and issues for performing these types of computations on parallel systems. We will describe system design issues (memory, I/O, mass storage, operating system requirements, special user interface issues, interconnects, bandwidths, and programming languages) involved in parallel methods for scalable classical, semiclassical, quantum, molecular mechanics, and continuum models; molecular nanotechnology computer-aided designs (NanoCAD) techniques; visualization using virtual reality techniques of structural models and assembly sequences; software required to control mini robotic manipulators for positional control; scalable numerical algorithms for reliability, verifications and testability. There appears no fundamental obstacle to simulating molecular compilers and molecular computers on high performance parallel computers, just as the Boeing 777 was simulated on a computer before manufacturing it.

Saini, Subhash↗
DOE OSTI2023

Reinforcement Learning for Load-balanced Parallel Particle Tracing

We explore an online reinforcement learning (RL) paradigm to dynamically optimize parallel particle tracing performance in distributed-memory systems. Our method combines three novel components: (1) a work donation algorithm, (2) a high-order workload estimation model, and (3) a communication cost model. First, we design an RL-based work donation algorithm. Our algorithm monitors workloads of processes and creates RL agents to donate data blocks and particles from high-workload processes to low-workload processes to minimize program execution time. The agents learn the donation strategy on the fly based on reward and cost functions designed to consider processes' workload changes and data transfer costs of donation actions. Second, we propose a workload estimation model, helping RL agents estimate the workload distribution of processes in future computations. Third, we design a communication cost model that considers both block and particle data exchange costs, helping RL agents make effective decisions with minimized communication costs. We demonstrate that our algorithm adapts to different flow behaviors in large-scale fluid dynamics, ocean, and weather simulation data. Our algorithm improves parallel particle tracing performance in terms of parallel efficiency, load balance, and costs of I/O and communication for evaluations with up to 16,384 processors.

Distributed and parallel particle tracing↗
NASA NTRS1988

TRAPEDS: Producing traces for multicomputers via execution-driven simulation

Trace-driven simulation is an important aid in performance analysis of computer systems. Capturing address traces for these simulations is a difficult problem for single processors and particularly for multicomputers. Even when existing trace methods can be used on multicomputers, the amount of collected data typically grows with the number of processors, so I/O and trace storage costs increase. A new technique is presented which modifies the executable code to dynamically collect the address trace from the user code and analyzes this trace during the execution of the program. This method helps resolve the I/O and storage problems and facilitates parallel analysis of the address trace. If a trace stored on disk is desired, the generated trace information can also be written to files during execution, with a resultant drop in program execution speed. An initial implementation on the Intel iPSC/2 hypercube multicomputer is detailed, and sample simulation results are presented. The effect of this trace collection method on execution time is illustrated.

Stunkel, Craig B.↗
NASA NTRS1989

TRAPEDS - Producing traces for multicomputers via execution driven simulation

Trace-driven simulation is an important aid in performance analysis of computer systems. Capturing address traces for these simulations is a difficult problem for single processors and particularly for multicomputers. Even when existing trace methods can be used on multicomputers, the amount of collected data typically grows with the number of processors, so I/O and trace storage costs increase. A new technique is presented which modifies the executable code to dynamically collect the address trace from the user code and analyzes this trace during the execution of the program. This method helps resolve the I/O and storage problems and facilitates parallel analysis of the address trace. If a trace stored on disk is desired, the generated trace information can also be written to files during execution, with a resultant drop in program execution speed. An initial implementation on the Intel iPSC/2 hypercube multicomputer is detailed, and sample simulation results are presented. The effect of this trace collection method on execution time is illustrated.

Stunkel, Craig B.↗
DOE OSTI2025

Direct I/O for RNTuple Columnar Data

RNTuple is the new columnar data format designed as the successor to ROOT’s TTree format. It allows to make use of modern hardware capabilities and is expected to be used in production by the LHC experiments during the HL-LHC. In this paper, we discuss the usage of Direct I/O to fully exploit modern SSDs, especially in the context of the recent addition of parallel RNTuple writing. We describe the alignment requirements imposed by Direct I/O and approaches to meet them for columnar data formats. Finally, we discuss performance results for both writing and reading, in synthetic benchmarks as well as real-world applications.

Hahnfeld, Jonas [CERN; Goethe U., Frankfurt (main)↗
DOE OSTI2025

Leveraging Pre-Built Catalogs and Object-Level Scheduling to Eliminate I/O Bottlenecks in HPC Environments

Modern High-Performance Computing (HPC) environments face mounting challenges due to the shift from large to small file datasets, along with an increasing number of users and parallelized applications. As HPC systems rely on Parallel File Systems (PFS), such as Lustre for data processing, performance bottlenecks stemming from Object Storage Target (OST) contention have become a significant concern. Existing solutions, such as LADS with its object-level scheduling approach, fall short in large-scale HPC environments due to their inability to effectively address metadata I/O bottlenecks and the growing number of I/O processes. This study highlights the pressing need for a comprehensive solution that tackles both OST contention and metadata I/O challenges in diverse HPC workloads. To address these challenges, we propose SwiftLoad, an object-level I/O scheduling framework that leverages a metadata catalog to enhance the performance and efficiency of parallel HPC utilities. The adoption of the metadata catalog mitigates the metadata I/O bottlenecks that commonly occur in HPC utilities, a challenge that is particularly pronounced in object-level I/O scheduling. SwiftLoad addresses OST contention and the uneven distribution of I/O processes across different OSTs through mathematical modeling and incorporates a Loader Configuration Module to regulate the number of I/O processes. Evaluated with two representative utilities—data deduplication profiling and data augmentation—SwiftLoad achieved performance improvements of up to 5.63x and 11.0x, respectively, on a production supercomputer.

HPC↗
DOE OSTI2021

Improving I/O Performance for Exascale Applications through Online Data Layout Reorganization

The applications being developed within the U.S. Exascale Computing Project (ECP) to run on imminent Exascale computers will generate scientific results with unprecedented fidelity and record turn-around time. Many of these codes are based on particle-mesh methods and use advanced algorithms, especially dynamic load-balancing and mesh-refinement, to achieve high performance on Exascale machines. Yet, as such algorithms improve parallel application efficiency, they raise new challenges for I/O logic due to their irregular and dynamic data distributions. Thus, while the enormous data rates of Exascale simulations already challenge existing file system write strategies, the need for efficient read and processing of generated data introduces additional constraints on the data layout strategies that can be used when writing data to secondary storage. We review these I/O challenges and introduce two online data layout reorganization approaches for achieving good tradeoffs between read and write performance. We demonstrate the benefits of using these two approaches for the ECP particle-in-cell simulation WarpX, which serves as a motif for a large class of important Exascale applications. Here, we show that by understanding application I/O patterns and carefully designing data layouts we can increase read performance by more than 80 percent.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Evolution of the ROOT Tree I/O

The ROOT TTree data format encodes hundreds of petabytes of High Energy and Nuclear Physics events. Its columnar layout drives rapid analyses, as only those parts (“branches”) that are really used in a given analysis need to be read from storage. Its unique feature is the seamless C++ integration, which allows users to directly store their event classes without explicitly defining data schemas. In this contribution, we present the status and plans of the future ROOT 7 event I/O. Along with the ROOT 7 interface modernization, we aim for robust, where possible compile-time safe C++ interfaces to read and write event data. On the performance side, we show first benchmarks using ROOT’s new experimental I/O subsystem that combines the best of TTrees with recent advances in columnar data formats. A core ingredient is a strong separation of the high-level logical data layout (C++ classes) from the low-level physical data layout (storage backed nested vectors of simple types). We show how the new, optimized physical data layout speeds up serialization and deserialization and facilitates parallel, vectorized and bulk operations. This lets ROOT I/O run optimally on the upcoming ultra-fast NVRAM storage devices, as well as file-less storage systems such as object stores.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
NASA NTRS1991

Collecting address traces from parallel computers

Trace driven simulation is a well-established method of performance analysis for single processor computer systems. However, efficient and accurate memory address tracing for parallel computer systems is not well understood. In this paper we present a critical survey of recently implemented approaches to address tracing and highlight the issues specific to collection of traces for both shared and distributed memory parallel computers. These issues include potential distortion of the relative ordering of events by the address tracing activity, realistic interleaving of addresses generated by multiple processors, and I/O and storage problems associated with collecting traces for large parallel systems. The strengths and weaknesses of the parallel tracing approaches are described.

Stunkel, Craig B.↗