Search NASA⌕ Search

SEARCH · Search NASA

Results for “storage cache”

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 73 records · Page 4

FPDeep: Scalable Acceleration of CNN Training on Deeply-Pipelined FPGA Clusters

In this paper, we propose a framework called FPDeep, which uses a hybrid of model and layer paral- lelism to configure distributed reconfigurable clusters to train DNNs. This approach has numerous benefits. First, the design does not suffer from batch size growth. Second, novel workload and weight partitioning leads to balanced loads of both among nodes. And third, the entire system is fine-grained pipeline. This leads to high parallelism and utilization and also minimizes the time features need to be cached while waiting for back-propagation.

Wang, Tianqi↗

Efficient multitasking of Choleski matrix factorization on CRAY supercomputers

A Choleski method is described and used to solve linear systems of equations that arise in large scale structural analysis. The method uses a novel variable-band storage scheme and is structured to exploit fast local memory caches while minimizing data access delays between main memory and vector registers. Several parallel implementations of this method are described for the CRAY-2 and CRAY Y-MP computers demonstrating the use of microtasking and autotasking directives. A portable parallel language, FORCE, is used for comparison with the microtasked and autotasked implementations. Results are presented comparing the matrix factorization times for three representative structural analysis problems from runs made in both dedicated and multi-user modes on both computers. CPU and wall clock timings are given for the parallel implementations and are compared to single processor timings of the same algorithm.

Overman, Andrea L.↗

Parallel variable-band Choleski solvers for computational structural analysis applications on vector multiprocessor supercomputers

A Choleski method used to solve linear systems of equations that arise in large scale structural analyses is described. The method uses a novel variable-band storage scheme and is structured to exploit fast local memory caches while minimizing data access delays between main memory and vector registers. Several parallel implementations of this method are described for the CRAY-2 and CRAY Y-MP computers demonstrating the use of microtasking and autotasking directives. A portable parallel language, FORCE, is also used for two different parallel implementations, demonstrating the use of CRAY macrotasking. Results are presented comparing the matrix factorization times for three representative structural analysis problems from runs made in both dedicated and multi-user modes on both the CRAY-2 and CRAY Y-MP computers. CPU and wall clock timings are given for the various parallel methods and are compared to single processor timings of the same algorithm. Computation rates over 1 GIGAFLOP (1 billion floating point operations per second) on a four processor CRAY-2 and over 2 GIGAFLOPS on an eight processor CRAY Y-MP are demonstrated as measured by wall clock time in a dedicated environment. Reduced wall clock times for the parallel methods relative to the single processor implementation of the same Choleski algorithm are also demonstrated for runs made in multi-user mode.

Poole, E. L.↗

Motivation and Design of the Sirocco Storage System Version 1.0.

Sirocco is a massively parallel, high performance storage system for the exascale era. It emphasizes client-to-client coordination, low server-side coupling, and free data movement to improve resilience and performance. Its architecture is inspired by peer-to-peer and victim- cache architectures. By leveraging these ideas, Sirocco natively supports several media types, including RAM, flash, disk, and archival storage, with automatic migration between levels. Sirocco also includes storage interfaces and support that are more advanced than typical block storage. Sirocco enables clients to efficiently use key-value storage or block-based storage with the same interface. It also provides several levels of transactional data updates within a single storage command, including full ACID-compliant updates. This transaction support extends to updating several objects within a single transaction. Further support is provided for con- currency control, enabling greater performance for workloads while providing safe concurrent modification. By pioneering these and other technologies and techniques in the storage system, Sirocco is poised to fulfill a need for a massively scalable, write-optimized storage system for exascale systems. This is version 1.0 of a document reflecting the current and planned state of Sirocco. Further versions of this document will be accessible at http://www.cs.sandia.gov/Scalable_IO/ sirocco .

97 MATHEMATICS AND COMPUTING↗

A distributed parallel storage architecture and its potential application within EOSDIS

We describe the architecture, implementation, use of a scalable, high performance, distributed-parallel data storage system developed in the ARPA funded MAGIC gigabit testbed. A collection of wide area distributed disk servers operate in parallel to provide logical block level access to large data sets. Operated primarily as a network-based cache, the architecture supports cooperation among independently owned resources to provide fast, large-scale, on-demand storage to support data handling, simulation, and computation.

Johnston, William E.↗

I/O performance studies of analysis workloads on production and dedicated resources at CERN

The recent evolutions of the analysis frameworks and physics data formats of the LHC experiments provide the opportunity of using central analysis facilities with a strong focus on interactivity and short turnaround times, to complement the more common distributed analysis on the Grid. In order to plan for such facilities, it is essential to know in detail the performance of the combination of a given analysis framework, of a specific analysis and of the installed computing and storage resources. This contribution describes performance studies performed at CERN, using the EOS disk-based storage, either directly or through an XCache instance, from both batch resources and highperformance compute nodes which could be used to build an analysis facility. A variety of benchmarks, both synthetic and based on real-world physics analyses and their corresponding input datasets, are utilized. In particular, the RNTuple format from the ROOT project is put to the test and compared to the latest version of the TTree format, and the impact of caches is assessed. In addition, we assessed the difference in performance between the use of storage system specific protocols, like XRootd, and FUSE. The results of this study are intended to be a valuable input in the design of analysis facilities, at CERN and elsewhere.

Sciabà, Andrea↗

Collective I/O Tuning Using Analytical and Machine-Learning Models

The ever larger demand of scientific applications for computation and data is currently driving a continuous increase in scale of parallel computers. The inherent complexity of scaling up a computing systems in terms of both hardware and software stack exposes an increasing number of factors impacting the performance and complicating the process of optimization. In particular, the optimization of parallel I/O has become increasingly challenging due to increasing storage hierarchy and well known performance variability of shared storage systems. This paper focuses on model-based autotuning of the two-phase collective I/O algorithm from a popular MPI distribution on the Blue Gene/Q architecture. We propose a novel hybrid model, constructed as a composition of analytical models for communication and storage operations and black-box models for the performance of the individual operations. We perform an in-depth study of the complexity involved in performance modeling including architecture, software stack and noise. In particular we address this challenges of modeling the performance of shared storage systems by building a benchmark that helps synthesizing factors such as topology, file caching, and noise. The experimental results show that the hybrid approach produces significantly better results than state-of-the-art machine learning approaches and shows a higher robustness to noise, at the cost of a higher modeling complexity

I/O performance modeling↗

Performance of hashed cache data migration schemes on multicomputers

After conducting an examination of several data-migration mechanisms which permit an explicit and controlled mapping of data to memory, a set of schemes for storage and retrieval of off-processor array elements is experimentally evaluated and modeled. All schemes considered have their basis in the use of hash tables for efficient access of nonlocal data. The techniques in question are those of hashed cache, partial enumeration, and full enumeration; in these, nonlocal data are stored in hash tables, so that the operative difference lies in the amount of memory used by each scheme and in the retrieval mechanism used for nonlocal data.

Hiranandani, Seema↗

Sparse triangular solves for ILU revisited: data layout crucial to better performance.; .

A key to good processor utilization for sparse matrix computations is storing the data in the format that is most conducive to fast access by the memory system. In particular, for sparse matrix triangular solves the traditional compressed sparse matrix format is poor, and minor adjustments to the data structure can increase the processor utilization dramatically. Such adjustments involve storing the L and U factors separately and storing the U rows 'backwards' so that they are accessed in a simple streaming fashion during the triangular solves. Changes to the PETSc libraries to use this modified storage format resulted in over twice the floating-point rate for some matrices. This improvement can be accounted for by a decrease in the cache misses and TLB (transaction lookaside buffer) misses in the modified code.

99 GENERAL AND MISCELLANEOUS↗

DDStore: Distributed Data Store for Scalable Training of Graph Neural Networks on Large Atomistic Modeling Datasets

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.

Choi, Jong Youl↗

HybridPlan: A Capacity Planning Technique for Projecting Storage Requirements in Hybrid Storage Systems

Economic forces, driven by the desire to introduce flash into the high-end storage market without changing existing software-base, have resulted in the emergence of solid-state drives (SSDs), flash packaged in HDD form factors and capable of working with device drivers and I/O buses designed for HDDs. Unlike the use of DRAM for caching or buffering, however, certain idiosyncrasies of NAND Flash-based solid-state drives (SSDs) make their integration into hard disk drive (HDD)-based storage systems nontrivial. Flash memory suffers from limits on its reliability, is an order of magnitude more expensive than the magnetic hard disk drives (HDDs), and can sometimes be as slow as the HDD (due to excessive garbage collection (GC) induced by high intensity of random writes). Given the complementary properties of HDDs and SSDs in terms of cost, performance, and lifetime, the current consensus among several storage experts is to view SSDs not as a replacement for HDD, but rather as a complementary device within the high-performance storage hierarchy. Thus, we design and evaluate such a hybrid storage system with HybridPlan that is an improved capacity planning technique to administrators with the overall goal of operating within cost-budgets. HybridPlan is able to find the most cost-effective hybrid storage configuration with different types of SSDs and HDDs

Capacity Planning↗

ATLAS Data Carousel

The ATLAS experiment at CERN's LHC stores detector and simulation data in raw and derived data formats across more than 150 Grid sites world-wide, currently in total about 200PB on disk and 250PB on tape. Data have different access characteristics due to various computational workflows, and can be accessed from different media, such as remote I/O, disk cache on hard disk drives or SSDs. Also, larger data centers provide the majority of offline storage capability via tape systems. For the High-Luminosity LHC (HL-LHC), the estimated data storage requirements are several factors bigger than the present forecast of available resources, based on a flat budget assumption. On the computing side, ATLAS Distributed Computing was very successful in the last years with high performance and high throughput computing integration and in using opportunistic computing resources for the Monte Carlo simulation. On the other hand, equivalent opportunistic storage does not exist. ATLAS started the "Data Carousel" project to increase the usage of less expensive storage, i.e. tapes or even commercial storage, so it is not limited to tape technologies exclusively. Data Carousel orchestrates data processing between workload management, data management, and storage services with the bulk data resident on offline storage. The processing is executed by staging and promptly processing a sliding window of inputs onto faster buffer storage, such that only a small percentage of input data are available at any one time. With this project, we aim to demonstrate that this is the natural way to dramatically reduce our storage cost. The first phase of the project was started in the fall of 2018 and was related to I/O tests of the sites archiving systems. Phase II now requires a tight integration of the workload and data management systems. Additionally, the Data Carousel studies the feasibility to run multiple computing workflows from tape. The project is progressing very well and the results presented in this document will be used before the LHC Run 3.

97 MATHEMATICS AND COMPUTING↗

Propagation Pattern for Moment Representation of the Lattice Boltzmann Method

A propagation pattern for the moment representation of the regularized lattice Boltzmann method (LBM) in three dimensions is presented. Using effectively lossless compression, the simulation state is stored as a set of moments of the lattice Boltzmann distribution function, instead of the distribution function itself. An efficient cache-aware propagation pattern for this moment representation has the effect of substantially reducing both the storage and memory bandwidth required for LBM simulations. This article extends recent work with the moment representation by expanding the performance analysis on central processing unit (CPU) architectures, considering how boundary conditions are implemented, and demonstrating the effectiveness of the moment representation on a graphics processing unit (GPU) architecture.

42 ENGINEERING↗

KG-Hub—building and exchanging biological knowledge graphs

Knowledge graphs (KGs) are a powerful approach for integrating heterogeneous data and making inferences in biology and many other domains, but a coherent solution for constructing, exchanging, and facilitating the downstream use of KGs is lacking. Here we present KG-Hub, a platform that enables standardized construction, exchange, and reuse of KGs. Features include a simple, modular extract–transform–load pattern for producing graphs compliant with Biolink Model (a high-level data model for standardizing biological data), easy integration of any OBO (Open Biological and Biomedical Ontologies) ontology, cached downloads of upstream data sources, versioned and automatically updated builds with stable URLs, web-browsable storage of KG artifacts on cloud infrastructure, and easy reuse of transformed subgraphs across projects. Current KG-Hub projects span use cases including COVID-19 research, drug repurposing, microbial–environmental interactions, and rare disease research. KG-Hub is equipped with tooling to easily analyze and manipulate KGs. KG-Hub is also tightly integrated with graph machine learning (ML) tools which allow automated graph ML, including node embeddings and training of models for link prediction and node classification.

59 BASIC BIOLOGICAL SCIENCES↗

Performance of redundant disk array organizations in transaction processing environments

A performance evaluation is conducted for two redundant disk-array organizations in a transaction-processing environment, relative to the performance of both mirrored disk organizations and organizations using neither striping nor redundancy. The proposed parity-striping alternative to striping with rotated parity is shown to furnish rapid recovery from failure at the same low storage cost without interleaving the data over multiple disks. Both noncached systems and systems using a nonvolatile cache as the controller are considered.

Mourad, Antoine N.↗

Horizontally scaling dChache SRM with the Terracotta platform

The dCache disk caching file system has been chosen by a majority of LHC experiments Tier 1 centers for their data storage needs. It is also deployed at many Tier 2 centers. The Storage Resource Manager (SRM) is a standardized grid storage interface and a single point of remote entry into dCache, and hence is a critical component. SRM must scale to increasing transaction rates and remain resilient against changing usage patterns. The initial implementation of the SRM service in dCache suffered from an inability to support clustered deployment, and its performance was limited by the hardware of a single node. Using the Terracotta platform, we added the ability to horizontally scale the dCache SRM service to run on multiple nodes in a cluster configuration, coupled with network load balancing. This gives site administrators the ability to increase the performance and reliability of SRM service to face the ever-increasing requirements of LHC data handling. In this paper we will describe the previous limitations of the architecture SRM server and how the Terracotta platform allowed us to readily convert single node service into a highly scalable clustered application.

99 GENERAL AND MISCELLANEOUS↗

Striped Data Analysis Framework

Traditionally, High Energy data analysis is based on the model where data are stored in files and analyzed by running multiple analysis processes, each reading one or more of the data files. This process involves repeated data reduction step, which produces smaller files, which is time consuming and leads to data duplication. We propose an alternative approach to data storage and analysis, based on the Big Data technologies. The idea is to store each element of data once and only once in a distributed scalable database and analyze data by reading only "interesting" pieces of data from the database. To make this approach possible, we developed columnar Striped Data Representation Format as the basis of the framework. Traditional columnar approach allows for efficient analysis of complex data structures. While keeping all the benefits of columnar data representation, striped mechanism goes further by enabling efficient parallelization of computations and flexible dist ribution of data analysis. The framework includes scalable and elastic data storage, compute and user analysis backend components. The framework uses off-the shelf web services and data caching technologies as the compute/data co-location mechanism. Flexible architecture allows the framework run in the cloud using container technologies. The framework offers Python/Jupyter as the user analysis backend platform, but can also run in command line or batch mode. In the article, we will present the results of the FNAL-LDRD-2016-032 FNAL LDRD project, the design, implementation, features and performance characteristics of the Striped Data Analysis Framework.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗