Quantifying I/O and Communication Traffic Interference on Dragonfly Networks Equipped with Burst Buffers
Explore the source record for details and available documents.
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.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
Abstract not provided.
Abstract not provided.
A computer-executable method, system, and computer program product for managing I/O requests from a compute node in communication with a data storage system, including a first burst buffer node and a second burst buffer node, the computer-executable method, system, and computer program product comprising striping data on the first burst buffer node and the second burst buffer node, wherein a first portion of the data is communicated to the first burst buffer node and a second portion of the data is communicated to the second burst buffer node, processing the first portion of the data at the first burst buffer node, and processing the second portion of the data at the second burst buffer node.
A computer-executable method, system, and computer program product for managing meta-data in a distributed storage system, wherein the distributed storage system includes one or more burst buffers enabled to operate with a distributed key-value store, the co computer-executable method, system, and computer program product comprising receiving a request for meta-data associated with a block of data stored in a first burst buffer of the one or more burst buffers in the distributed storage system, wherein the meta data is associated with a key-value, determining which of the one or more burst buffers stores the requested metadata, and upon determination that a first burst buffer of the one or more burst buffers stores the requested metadata, locating the key-value in a portion of the distributed key-value store accessible from the first burst buffer.
Solid-State Drive (SSD) burst buffer nodes are interposed into a parallel supercomputing cluster to enable fast burst checkpoint of cluster memory to or from nearby interconnected solid-state storage with asynchronous migration between the burst buffer nodes and slower more distant disk storage. The SSD nodes also perform tasks offloaded from the compute nodes or associated with the checkpoint data. For example, the data for the next job is preloaded in the SSD node and very fast uploaded to the respective compute node just before the next job starts. During a job, the SSD nodes perform fast visualization and statistical analysis upon the checkpoint data. The SSD nodes can also perform data reduction and encryption of the checkpoint data.
At present, I/O performance analysis requires different tools to characterize individual components of the I/O subsystem, and institutional I/O expertise is relied upon to translate these disparate data into an integrated view of application performance. This process is labor-intensive and not sustainable as the storage hierarchy deepens and system complexity increases. To address this growing disparity, we have developed the Total Knowledge of I/O (TOKIO) framework to combine the insights from existing component-level monitoring tools and provide a holistic view of performance across the entire I/O stack. A reference implementation of TOKIO, pytokio, is presented here. Using monitoring tools included with Cray XC and ClusterStor systems alongside commonly deployed community-supported tools, we demonstrate how pytokio provides a lightweight foundation for holistic I/O performance analyses on two Cray XC systems deployed at different HPC centers. We present results from integrated analyses that allow users to quantify the degree of I/O contention that affected their jobs and probabilistically identify unhealthy storage devices that impacted their performance.We also apply pytokio to inspect the utilization of NERSC’s DataWarp burst buffer and demonstrate how pytokio can be used to identify users and applications who may stand to benefit most from migrating their workloads from Lustre to the burst buffer.
A key task in astronomy is to locate astronomical objects in images and to characterize them according to physical parameters such as brightness, color, and morphology. This task, known as cataloging, is challenging for several reasons: many astronomical objects are much dimmer than the sky background, labeled data is generally unavailable, overlapping astronomical objects must be resolved collectively, and the datasets are enormous – terabytes now, petabytes soon. In this work, we infer an astronomical catalog from 55 TB of imaging data using Celeste, a Bayesian variational inference code written entirely in the high-productivity programming language Julia. Using over 1.3 million threads on 650,000 Intel Xeon Phi cores of the Cori Phase II supercomputer, Celeste achieves a peak rate of 1.54 DP PFLOP/s. Celeste is able to jointly optimize parameters for 188 M stars and galaxies, loading and processing 178 TB across 8192 nodes in 14.6 min. To achieve this, Celeste exploits parallelism at multiple levels (cluster, node, and thread) and accelerates I/O through Cori’s burst buffer. Julia’s native performance enables Celeste to employ high-level constructs without resorting to hand-written or generated low-level code (C/C++/Fortran) while still achieving petascale performance.
To solve the supercomputing challenge of memory keeping up with processing speed, a team at Los Alamos National Laboratory developed two innovative memory management and storage technologies. Burst buffers peel off data onto flash memory to support the checkpoint/restart paradigm of large simulations. MarFS adds a thin software layer enabling a new tier for campaign storage—based on inexpensive, failure-prone disk drives—between disk drives and tape archives.
Scientific exploration generates expanding volumes of data that commonly require High Performance Computing (HPC) systems to facilitate research. HPC systems are complex ecosystems of hardware and software that frequently are not user friendly. The Usable Data Abstractions (UDA) project set out to build usable software for scientific workflows in HPC environments by undertaking multiple rounds of qualitative user research. Qualitative research investigates how individuals accomplish their work and our interview-based study surfaced a variety of insights about the experiences of working in and with HPC ecosystems. This report examines multiple facets to the experiences of scientists and developers using and supporting HPC systems. We discuss how stakeholders grasp the design and configuration of these systems, the impacts of abstraction layers on their ability to successfully do work, and the varied perceptions of time that shape this work. Examining the adoption of the Cori HPC at NERSC we explore the anticipations and lived experiences of users interacting with this system’s novel storage feature, the Burst Buffer. We present lessons learned from across these insights to illustrate just some of the challenges HPC facilities and their stakeholders need to account for when procuring and supporting these essential scientific resources to ensure their usability and utility to a variety of scientific practices.
Data Elavator allows extensive optimizations, such as overlapping read and write operations, choosing I/O modes, and aligning buffer boundaries. Is as much as 4.2X faster than Cray DataWarp, the start-of-art software for burst buffer, and 4X faster than directly writing to PFS.
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.
Scientific applications at exascale generate and analyze massive amounts of data. A critical requirement of these applications is the capability to access and manage this data efficiently on exascale systems. As such, parallel I/O, the key technology enables moving data between compute nodes and storage, faces monumental challenges from new applications, memory, and storage architectures considered in the designs of exascale systems. As the storage hierarchy is expanding to include node-local persistent memory, burst buffers, etc., as well as disk-based storage, data movement among these layers must be efficient. Parallel I/O libraries of the future should be capable of handling file sizes of many terabytes and beyond. In this paper, we describe new capabilities we have developed in Hierarchical Data Format version 5 (HDF5), the most popular parallel I/O library for scientific applications. HDF5 is one of the most used libraries at the leadership computing facilities for performing parallel I/O on existing HPC systems. The state-of-the-art features we describe include: Virtual Object Layer (VOL), Data Elevator, asynchronous I/O, full-featured single-writer and multiple-reader (Full SWMR), and parallel querying. In this paper, we introduce these features, their implementations, and the performance and feature benefits to applications and other libraries.
As the exascale computing age emerges, data related issues are becoming critical factors that determine how and where we do computing. Popular approaches used by traditional I/O solution and storage libraries become increasingly bottlenecked due to their assumptions about data movement, re-organization, and storage. While, new technologies, such as "burst buffers", can help address some of the short-term performance issues, it is essential that we reexamine the underlying storage and I/O infrastructure to effectively support requirements and challenges at exascale and beyond. In this paper we present a new approach to the exascale Storage System and I/O (SSIO), which is based on allowing users to inject application knowledge into the system and leverage this knowledge to better manage, store, and access large data volumes so as to minimize the time to scientific insights. Central to our approach is the distinction between the data, metadata, and the knowledge contained therein, transferred from the user to the system by describing "utility" of data as it ages.
FanStore is a shared object store that enables efficient and scalable neural network training on supercomputers. By providing a global cache layer on node-local burst buffers using a compressed representation, it significantly enhances the processing capability of deep learning (DL) applications on existing hardware. In addition, FanStore allows POSIX-compliant file access to the compressed data in user space. We investigate the tradeoff between runtime overhead and data compression ratio using real-world datasets and applications, and propose a compressor selection algorithm to maximize storage capacity given performance constraints. We consider both asynchronous (i.e., with prefetching) and synchronous I/O strategies, and propose mechanisms for selecting compressors for both approaches. Using FanStore, the same storage hardware can host 2-13x more data for example applications without significant runtime overhead. Empirically, our experiments show that FanStore scales to 512 compute nodes with near linear performance scalability.