GLANCED-IO: Taming I/O Optimization for Deep Learning at Scale
Explore the source record for details and available documents.
Computer science↗
SEARCH · Engineering Papers
Results for “I/O Optimization”
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.
Input/output behavior of supercomputing applications
The collection and analysis of supercomputer I/O traces and their use in a collection of buffering and caching simulations are described. This serves two purposes. First, it gives a model of how individual applications running on supercomputers request file system I/O, allowing system designer to optimize I/O hardware and file system algorithms to that model. Second, the buffering simulations show what resources are needed to maximize the CPU utilization of a supercomputer given a very bursty I/O request rate. By using read-ahead and write-behind in a large solid stated disk, one or two applications were sufficient to fully utilize a Cray Y-MP CPU.
Miller, Ethan L.↗
I/O in Machine Learning Applications on HPC Systems: A 360-degree Survey
Growing interest in Artificial Intelligence (AI) has resulted in a surge in demand for faster methods of Machine Learning (ML) model training and inference. This demand for speed has prompted the use of high performance computing (HPC) systems that excel in managing distributed workloads. Because data is the main fuel for AI applications, the performance of the storage and I/O subsystem of HPC systems is critical. In the past, HPC applications accessed large portions of data written by simulations or experiments or ingested data for visualizations or analysis tasks. ML workloads perform small reads spread across a large number of random files. This shift of I/O access patterns poses several challenges to modern parallel storage systems. In this paper, we survey I/O in ML applications on HPC systems, and target literature within a 6-year time window from 2019 to 2024. We define the scope of the survey, provide an overview of the common phases of ML, review available profilers and benchmarks, examine the I/O patterns encountered during offline data preparation, training, and inference, and explore I/O optimizations utilized in modern ML frameworks and proposed in recent literature. Lastly, we seek to expose research gaps that could spawn further R&D.
97 MATHEMATICS AND COMPUTING↗
Methodology and Application of HPC I/O Characterization with MPIProf and IOT
Combining the strengths of MPIProf and IOT, an efficient and systematic method is devised for I/O characterization at the per-job, per-rank, per-file and per-call levels of HPC programs running on the NASA Advanced Supercomputing Center. This method is applied to answer four I/O questions in this paper. A total of 13 MPI programs and 15 cases, ranging from 24 to 5968 ranks, are analyzed to establish the I/O landscape from answers to the four questions. Four of the 13 programs use MPI I/O and the behavior of their collective writes depends on the specific implementation of the MPI library used. The SGI MPT library, the prevailing MPI library for our systems, was found to gather small writes from a large number of ranks to perform larger writes by a small subset of collective buffering ranks. The number of collective buffering ranks invoked by MPT depends on the Lustre stripe count and the number of nodes used for the run. A demonstration of varying the stripe count to achieve double-digit speedup of one program's I/O was presented. Another program, which concurrently opens private files by all ranks and could potentially create a heavy load on the Lustre servers, was identified. The ability to systematically characterize I/O for a large number of programs running on a supercomputer, seek I/O optimization opportunity and identify programs that could cause a high load and instability on the filesystems is important for pursuing exascale in a real production environment.
Characterization↗
I/O Performance Characterization of Lustre and NASA Applications on Pleiades
In this paper we study the performance of the Lustre file system using five scientific and engineering applications representative of NASA workload on large-scale supercomputing systems such as NASA s Pleiades. In order to facilitate the collection of Lustre performance metrics, we have developed a software tool that exports a wide variety of client and server-side metrics using SGI's Performance Co-Pilot (PCP), and generates a human readable report on key metrics at the end of a batch job. These performance metrics are (a) amount of data read and written, (b) number of files opened and closed, and (c) remote procedure call (RPC) size distribution (4 KB to 1024 KB, in powers of 2) for I/O operations. RPC size distribution measures the efficiency of the Lustre client and can pinpoint problems such as small write sizes, disk fragmentation, etc. These extracted statistics are useful in determining the I/O pattern of the application and can assist in identifying possible improvements for users applications. Information on the number of file operations enables a scientist to optimize the I/O performance of their applications. Amount of I/O data helps users choose the optimal stripe size and stripe count to enhance I/O performance. In this paper, we demonstrate the usefulness of this tool on Pleiades for five production quality NASA scientific and engineering applications. We compare the latency of read and write operations under Lustre to that with NFS by tracing system calls and signals. We also investigate the read and write policies and study the effect of page cache size on I/O operations. We examine the performance impact of Lustre stripe size and stripe count along with performance evaluation of file per process and single shared file accessed by all the processes for NASA workload using parameterized IOR benchmark.
Saini, Subhash↗
What to Support When You’re Compressing
Over the last nearly 20 years, lossy compression has become an essential aspect of HPC applications’ data pipelines, allowing them to overcome limitations in storage capacity and bandwidth and, in some cases, increase computational throughput and capacity. However, with the adoption of lossy compression comes the requirement to assess and control the impact lossy compression has on scientific outcomes. In this work, we take a major step forward in describing the state of practice and by characterizing workloads. We examine applications’ needs and compressors’ capabilities across 9 different supercomputing application domains. We present 24 takeaways that provide best practices for applications, operational impacts for facilities achieving compressed data, and gaps in application needs not addressed by production compressors that point towards opportunities for future compression research.
Error-Bounded Lossy Compression↗
SmartIO: An End-to-EndWorkflow for Runtime I/O Prediction and Cross-Layer Optimization of HPC Systems
Explore the source record for details and available documents.
Computer science↗
Enabling kilometer-scale E3SM land model simulation over North America: A new integrated framework solution
This study introduces a novel framework designed to enhance the performance, scalability, and portability of the kilometer-scale E3SM Land Model (km-ELM) within the E3SM modeling infrastructure. By seamlessly integrating cutting-edge data tools, we address existing challenges such as slow performance, limited scalability, and difficulties in software integration in current data-driven ELM simulation over large geographic areas. Our innovative approach leverages the KiloCraft data toolkit to generate unified inputs for simulations ranging from a single-cite case, to a 72,083-cell regional case to a continental configuration encompassing 21.6 million land grid cells at a 1 km × 1 km resolution. We conduct extensive strong- and weak-scaling experiments on three state-of-the-art supercomputers, utilizing up to 100,800 CPU cores across 2400 compute nodes to evaluate end-to-end metrics including wall-clock time, simulation-years-per-day (SYPD), initialization costs, and I/O throughput. Our results reveal the land (LND) component’s efficient scaling, demonstrating near-ideal weak scaling and strong-scaling parallel efficiencies reaching up to 87% at 50,400 cores. We confirm portability and reproducibility through bitwise-equivalent outputs across different machines using identical inputs over supported machines. Notably, at extreme scales, we identify I/O as a critical bottleneck and that leads to effective solution with the SCORPIO/ADIOS stack. Collectively, these findings validate the deployment of km-ELM at a continental scale with high parallel efficiency and provide essential guidance on configuration, decomposition, and I/O settings for optimized kilometer-scale land simulations in E3SM. This work emphasizes the innovative design and practical solutions that enhance the operational capabilities of km-ELM, focusing on software performance and scalability while leaving detailed scientific evaluations of simulated land processes for future investigations.
E3SM land model (ELM), km-ELM, scalability, perfor↗
Slave finite element for non-linear analysis of engine structures. Volume 2: Programmer's manual and user's manual
The programming aspects of SFENES are described in the User's Manual. The information presented is provided for the installation programmer. It is sufficient to fully describe the general program logic and required peripheral storage. All element generated data is stored externally to reduce required memory allocation. A separate section is devoted to the description of these files thereby permitting the optimization of Input/Output (I/O) time through efficient buffer descriptions. Individual subroutine descriptions are presented along with the complete Fortran source listings. A short description of the major control, computation, and I/O phases is included to aid in obtaining an overall familiarity with the program's components. Finally, a discussion of the suggested overlay structure which allows the program to execute with a reasonable amount of memory allocation is presented.
Witkop, D. L.↗
Improving I/O-aware Workflow Scheduling via Data Flow Characterization and trade-off Analysis
The scientific computing paradigm has transitioned from compute-intensive to I/O-intensive and memory-intensive in the past decade, especially when data-driven science has become common practice. Numerous empirical I/O-aware scheduling optimizations have been developed by incorporating I/O capacity and bandwidth as constraints into scheduling. Unfortunately, there is a lack of data flow (I/O) characterization tool and an understanding of trade-offs between concurrency, locality, and I/O bandwidth. To bridge the gap, this work 1) presents a set of descriptors to characterize, organize, and visualize I/O profiles, including flow size, I/O bandwidth, and operation count, which group data flows by I/O types, tasks, and files; 2) proposes an I/O Roofline model-based trade-off analysis to find the optimal trade-off between flow operational intensity, concurrency, and flow performance. The I/O descriptors generate useful insights into complicated I/O behaviors, suggesting distinct concurrency, storage, and scheduling to be used by types, tasks, and files. The proposed trade-off analysis guides scheduling decisions that generate resource assignment with the best flow parallelism. We evaluate our I/O-aware scheduling methodology on a highly I/O-intensive workflow–1000 Genomes. The experimental results demonstrate speedups of up to 2.4× compared to the state-of-the- art methods.
Guo, Luanzheng [BATTELLE (PACIFIC NW LAB)]↗
RAID 7 disk array
Each RAID level reflects a different design architecture. Associated with each is a backdrop of imposed limitations, as well as possibilities which may be exploited within the architectural constraints of that level. There are three unique features that differentiate RAID 7 from all other levels. RAID 7 is asynchronous with respect to usage of I/O data paths. Each I/O drive (includes all data and one parity drives) as well as each host interface (there may be multiple host interfaces) has independent control and data paths. This means that each can be accessed completely, independently, of the other. This is facilitated by a separate device cache for each device/interface as well. RAID 7 is asynchronous with respect to device hierarchy and data bus utilization. Each drive and each interface is connected to a high speed data bus controlled by the embedded operating system to make independent transfers to and from central cache. RAID 7 is asynchronous with respect to the operation of an embedded real time process oriented operating system. This means that exclusive and independent of the host, or multiple host paths, the embedded OS manages all I/O transfers asynchronously across the data and parity drives. A key factor to consider is that of the RAID 7's ability to anticipate and match host I/O usage patterns. This yields the following benefits over RAID's built around micro-code based architectures. RAID 7 appears to the host as a normally connected Big Fast Disk (BFD). RAID 7 appears, from the perspective of the individual disk devices, to minimize the total number of accesses and optimize read/write transfer requests. RAID 7 smoothly integrates the random demands of independent users with the principles of spatial and temporal locality. This optimizes small, large, and time sequenced I/O requests which results in users having an I/O performance which approaches performance to that of main memory.
Stout, Lloyd↗
High-Rate Data-Capture for an Airborne Lidar System
A high-rate data system was required to capture the data for an airborne lidar system. A data system was developed that achieved up to 22 million (64-bit) events per second sustained data rate (1408 million bits per second), as well as short bursts (less than 4 s) at higher rates. All hardware used for the system was off the shelf, but carefully selected to achieve these rates. The system was used to capture laser fire, single-photon detection, and GPS data for the Slope Imaging Multi-polarization Photo-counting Lidar (SIMPL). However, the system has applications for other laser altimeter systems (waveform-recording), mass spectroscopy, xray radiometry imaging, high-background- rate ranging lidar, and other similar areas where very high-speed data capture is needed. The data capture software was used for the SIMPL instrument that employs a micropulse, single-photon ranging measurement approach and has 16 data channels. The detected single photons are from two sources those reflected from the target and solar background photons. The instrument is non-gated, so background photons are acquired for a range window of 13 km and can comprise many times the number of target photons. The highest background rate occurs when the atmosphere is clear, the Sun is high, and the target is a highly reflective surface such as snow. Under these conditions, the total data rate for the 16 channels combined is expected to be approximately 22 million events per second. For each photon detection event, the data capture software reads the relative time of receipt, with respect to a one-per-second absolute time pulse from a GPS receiver, from an event timer card with 0.1-ns precision, and records that information to a RAID (Redundant Array of Independent Disks) storage device. The relative time of laser pulse firings must also be read and recorded with the same precision. Each of the four event timer cards handles the throughput from four of the channels. For each detection event, a flag is recorded that indicates the source channel. To accommodate the expected maximum count rate and also handle the other extreme of very low rates occurring during nighttime operations, the software requests a set amount of data from each of the event timer cards and buffers the data. The software notes if any of the cards did not return all the data requested and then accommodates that lower rate. The data is buffered to minimize the I/O overhead of writing the data to storage. Care was taken to optimize the reads from the cards, the speed of the I/O bus, and RAID configuration.
Valett, Susan↗
Optimizing Management of Persistent Data Structures in High-Performance Analytics
Large-scale data analytics workflows ingest massive input data into various data structures, including graphs and key-value datastores. These data structures undergo multiple transformations and computations and are typically reused in incremental and iterative analytics workflows. Persisting in-memory views of these data structures enables reusing them beyond the scope of a single program run while avoiding repetitive raw data ingestion overheads. Memory-mapped I/O enables persisting in-memory data structures without data serialization and deserialization overheads. However, memory-mapped I/O lacks the key feature of persisting consistent snapshots of these data structures for incremental ingestion and processing. The obstacles to efficient virtual memory snapshots using memory-mapped I/O include background writebacks outside the application’s control, and the significantly high storage footprint of such snapshots. To address these limitations, we present Privateer, a memory and storage management tool that enables storage-efficient virtual memory snapshotting while also optimizing snapshot I/O performance. Here, we integrated Privateer into Metall, a state-of-the-art persistent memory allocator for C++, and the Lightning Memory-Mapped Database (LMDB), a widely-used key-value datastore in data analytics and machine learning. Privateer optimized application performance by 1.22× when storing data structure snapshots to node-local storage, and up to 16.7× when storing snapshots to a parallel file system. Privateer also optimizes storage efficiency of incremental data structure snapshots by up to 11× using data deduplication and compression.
Computer science↗
Architecture and performance of Perlmutter's 35 PB ClusterStor E1000 all-flash file system
NERSC's newest system, Perlmutter, features a 35 PB all-flash Lustre file system built on HPE Cray ClusterStor E1000. Here, we present its architecture, early performance figures, and performance considerations unique to this architecture. We demonstrate the performance of E1000 OSSes through low-level Lustre tests that achieve over 90% of the theoretical bandwidth of the SSDs at the OST and LNet levels. We also show end-to-end performance for both traditional dimensions of I/O performance (peak bulk-synchronous bandwidth) and nonoptimal workloads endemic to production computing (small, incoherent I/Os at random offsets) and compare them to NERSC's previous system, Cori, to illustrate that Perlmutter achieves the performance of a burst buffer and the resilience of a scratch file system. Finally, we discuss performance considerations unique to all-flash Lustre and present ways in which users and HPC facilities can adjust their I/O patterns and operations to make optimal use of such architectures.
97 MATHEMATICS AND COMPUTING↗
Dual Channel Dual Staging: Hierarchical and Portable Staging for GPU-Based In-Situ Workflow
In-situ workflows have emerged as an attractive approach for addressing data movement challenges at very large scales. Since GPU-based architectures dominate the HPC landscapes, porting these in-situ workflows, and, specifically, the inter-application data exchange, to GPU-based systems can be challenging. Technologies such as GPUDirect RDMA (GDR), which is typically used for I/O in GPU applications as an optimization that circumvents the CPU overhead, can be leveraged to support bulk data exchanges between GPU applications. However, current GDR design often lacks performance portability across HPC clusters built with different hardware configurations. Furthermore, the local CPU may also be effectively used as an auxiliary communication mechanism to offload data exchanges. In this paper, we present a dual channel dual staging approach for efficient, scalable, and performance-portable inter-application data exchange for in-situ workflows. This approach exploits the data access pattern within in-situ workflows along with the inherent execution asynchrony to accelerate data exchanges and, at the same time, improve performance portability. Specifically, the dual channel dual staging method leverages both the local CPU and the remote data staging server to build a hierarchical joint staging area and uses this staging area to transform blocking inter-application bulk data exchanges into best-effort local data movements between GPU and CPU. The dual channel dual staging is implemented as a portability extension of the Dataspaces-GPU staging framework. We present an experimental evaluation of its performance, portability, and scalability using this implementation on three leadership GPU clusters. The evaluation results demonstrate that the dual channel dual staging method saves up to 75% in data-exchange time compared to host-based, GDR, and alternate portable designs, while maintaining scalability (up to 512 GPUs) and performance portability across the three platforms.
Zhang, Bo [University of Utah]↗
NASTRAN of the HEP
Preliminary plans for the conversion of NASTRAN to DENELCOR's heterogeneous element processor (HEP) are presented. First, a brief history of computer architecture and the architecture of the HEP computer system are discussed. Second, a description of the HEP computer system is provided. Lastly, the preliminary NASTRAN conversion plans for link management, I/O management, memory management, and code optimization are discussed.
Brown, W. K.↗
To Derive or Not to Derive: I/O Libraries Take Charge of Derived Quantities Computation
The ever-increasing volume of data produced by HPC simulations necessitates scalable methods for data exploration and knowledge extraction. Scientific data analysis often involves complex queries across distributed datasets, requiring manipulation of multiple primary variables and generating derived data that needs to be handled efficiently, creating challenges for applications that need to parse many large datasets. Relying on individual applications to handle all intermediate data generally leads to redundant computations across studies and unnecessary data transfers. In this paper, we investigate the performance of different approaches where applications define derived variables as quantities of interest (QoIs) and offload the computation and transfer of these QoIs to the I/O library. This significantly reduces redundancy and optimizes data movement across the distributed storage and processing infrastructure by allowing control over when and where derived variables are computed. We present a detailed analysis of the performance-storage trade-offs associated with different solutions and showcase results for our study on two large-scale datasets created from climate and combustion simulations.
Gainaru, Ana↗
Interactive Spectral Analysis and Computation (ISAAC)
Isaac is a task in the NSO external package for IRAF. A descendant of a FORTRAN program written to analyze data from a Fourier transform spectrometer, the current implementation has been generalized sufficiently to make it useful for general spectral analysis and other one dimensional data analysis tasks. The user interface for Isaac is implemented as an interpreted mini-language containing a powerful, programmable vector calculator. Built-in commands provide much of the functionality needed to produce accurate line lists from input spectra. These built-in functions include automated spectral line finding, least squares fitting of Voigt profiles to spectral lines including equality constraints, various filters including an optimal filter construction tool, continuum fitting, and various I/O functions.
Lytle, D. M.↗