Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Data parallelism”

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
DOE OSTI2025

GSplit: Scaling Graph Neural Network Training on Large Graphs via Split-Parallelism

Graph neural networks (GNNs), an emerging class of machine learning models for graphs, have gained popularity for their superior performance in various graph analytical tasks. Mini-batch training is commonly used to train GNNs on large graphs, and data parallelism is the standard approach to scale mini-batch training across multiple GPUs. Data parallel approaches contain redundant work as subgraphs sampled by different GPUs contain significant overlap. To address this issue, we introduce a hybrid parallel mini-batch training paradigm called Split parallelism. Split parallelism avoids redundant work by splitting the sampling, loading, and training of each mini-batch across multiple GPUs. Split parallelism, however, introduces communication overheads that can be more than the savings from removing redundant work. We further present a lightweight partitioning algorithm that probabilistically minimizes these overheads. We implement spllit parllelism in GSplit and show that it outperforms state-of-the-art mini-batch training systems like DGL, Quiver, and P3.

Lim, Seung-Hwan [ORNL] (ORCID:0000000194616866)↗
DOE OSTI2024

Data-flow parallelism for high-energy and nuclear physics computing frameworks

The processing tasks of a scientific workflow in high-energy and nuclear physics (HENP) can typically be represented as a directed acyclic graph formed according to the data flow—i.e. the data dependencies among algorithms executed as part of the workflow. With this representation, an HENP computing framework can optimally execute a workflow, exploiting the parallelism inherent among independent tasks. Despite such a natural description of a workflow, most HENP frameworks do not make use of technologies that provide concurrent execution of graph-based tasking structures. In this session, we describe Fermilab efforts to adopt a graph-based technology (specifically Intel’s oneTBB flow graph) for meeting the framework needs of its experiments, notably DUNE. After introducing the physics DUNE intends to explore, we will show that all common processing idioms supported by current HENP frameworks can naturally be supported by oneTBB’s data-flow technology, optimally leveraging the concurrent capabilities of the machine. In addition, we discuss collaborative efforts between Fermilab and the Intel oneTBB development team, who is considering improvements to the flow-graph technology to better support HENP use cases.

43 PARTICLE ACCELERATORS↗
DOE OSTI2024

Data-flow parallelism for high-energy and nuclear physics frameworks

The processing tasks of an event-processing workflow in high-energy and nuclear physics (HENP) can typically be represented as a directed acyclic graph formed according to the data flow—i.e. the data dependencies among algorithms executed as part of the workflow. With this representation, an HENP framework can optimally execute a workflow, exploiting the parallelism inherent among independent tasks. Despite such a natural description of a workflow, most HENP frameworks do not make use of technologies that provide concurrent execution of graph-based tasking structures. In this talk, we describe Fermilab efforts to adopt a graph-based technology (specifically Intel’s oneTBB flow graph) for meeting the framework needs of its experiments, notably DUNE. Building on the Meld project as presented at CHEP2023, we demonstrate that all common processing idioms supported by current frameworks can naturally be supported by oneTBB’s data-flow technology, optimally leveraging the concurrent capabilities of the machine. In addition, we discuss collaborative efforts between Fermilab and the Intel oneTBB development team, who is considering improvements to the flow-graph technology to better support HENP use cases.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
NASA NTRS1993

A data distributed parallel algorithm for ray-traced volume rendering

This paper presents a divide-and-conquer ray-traced volume rendering algorithm and a parallel image compositing method, along with their implementation and performance on the Connection Machine CM-5, and networked workstations. This algorithm distributes both the data and the computations to individual processing units to achieve fast, high-quality rendering of high-resolution data. The volume data, once distributed, is left intact. The processing nodes perform local ray tracing of their subvolume concurrently. No communication between processing units is needed during this locally ray-tracing process. A subimage is generated by each processing unit and the final image is obtained by compositing subimages in the proper order, which can be determined a priori. Test results on both the CM-5 and a group of networked workstations demonstrate the practicality of our rendering algorithm and compositing method.

Ma, Kwan-Liu↗
DOE OSTI2020

Time-resolved RIXS experiment with pulse-by-pulse parallel readout data collection using X-ray free electron laser

Abstract Time-resolved resonant inelastic X-ray scattering (RIXS) is one of the developing techniques enabled by the advent of X-ray free electron laser (FEL). It is important to evaluate how the FEL jitter, which is inherent in the self-amplified spontaneous emission process, influences the RIXS measurement. Here, we use a microchannel plate (MCP) based Timepix soft X-ray detector to conduct a time-resolved RIXS measurement at the Ti L 3 -edge on a charge-density-wave material TiSe 2 . The fast parallel Timepix readout and single photon sensitivity enable pulse-by-pulse data acquisition and analysis. Due to the FEL jitter, low detection efficiency of spectrometer, and low quantum yield of RIXS process, we find that less than 2% of the X-ray FEL pulses produce signals, preventing acquiring sufficient data statistics while maintaining temporal and energy resolution in this measurement. These limitations can be mitigated by using future X-ray FELs with high repetition rates, approaching MHz such as the European XFEL in Germany and LCLS-II in the USA, as well as by utilizing advanced detectors, such as the prototype used in this study.

36 MATERIALS SCIENCE↗
NASA NTRS1996

Parallel Climate Data Assimilation PSAS Package

We have designed and implemented a set of highly efficient and highly scalable algorithms for an unstructured computational package, the PSAS data assimilation package, as demonstrated by detailed performance analysis of systematic runs on up to 512node Intel Paragon. The equation solver achieves a sustained 18 Gflops performance. As the results, we achieved an unprecedented 100-fold solution time reduction on the Intel Paragon parallel platform over the Cray C90. This not only meets and exceeds the DAO time requirements, but also significantly enlarges the window of exploration in climate data assimilations.

PSAS data scalable algorithms Intel Paragon 512nod↗
NASA NTRS1994

High Performance FORTRAN

High performance FORTRAN is a set of extensions for FORTRAN 90 designed to allow specification of data parallel algorithms. The programmer annotates the program with distribution directives to specify the desired layout of data. The underlying programming model provides a global name space and a single thread of control. Explicitly parallel constructs allow the expression of fairly controlled forms of parallelism in particular data parallelism. Thus the code is specified in a high level portable manner with no explicit tasking or communication statements. The goal is to allow architecture specific compilers to generate efficient code for a wide variety of architectures including SIMD, MIMD shared and distributed memory machines.

Mehrotra, Piyush↗
DOE OSTI2022

Profiling the BLAST bioinformatics application for load balancing on high-performance computing clusters

Abstract Background The Basic Local Alignment Search Tool (BLAST) is a suite of commonly used algorithms for identifying matches between biological sequences. The user supplies a database file and query file of sequences for BLAST to find identical sequences between the two. The typical millions of database and query sequences make BLAST computationally challenging but also well suited for parallelization on high-performance computing clusters. The efficacy of parallelization depends on the data partitioning, where the optimal data partitioning relies on an accurate performance model. In previous studies, a BLAST job was sped up by 27 times by partitioning the database and query among thousands of processor nodes. However, the optimality of the partitioning method was not studied. Unlike BLAST performance models proposed in the literature that usually have problem size and hardware configuration as the only variables, the execution time of a BLAST job is a function of database size, query size, and hardware capability. In this work, the nucleotide BLAST application BLASTN was profiled using three methods: shell-level profiling with the Unix “time” command, code-level profiling with the built-in “profiler” module, and system-level profiling with the Unix “gprof” program. The runtimes were measured for six node types, using six different database files and 15 query files, on a heterogeneous HPC cluster with 500+ nodes. The empirical measurement data were fitted with quadratic functions to develop performance models that were used to guide the data parallelization for BLASTN jobs. Results Profiling results showed that BLASTN contains more than 34,500 different functions, but a single function, RunMTBySplitDB, takes 99.12% of the total runtime. Among its 53 child functions, five core functions were identified to make up 92.12% of the overall BLASTN runtime. Based on the performance models, static load balancing algorithms can be applied to the BLASTN input data to minimize the runtime of the longest job on an HPC cluster. Four test cases being run on homogeneous and heterogeneous clusters were tested. Experiment results showed that the runtime can be reduced by 81% on a homogeneous cluster and by 20% on a heterogeneous cluster by re-distributing the workload. Discussion Optimal data partitioning can improve BLASTN’s overall runtime 5.4-fold in comparison with dividing the database and query into the same number of fragments. The proposed methodology can be used in the other applications in the BLAST+ suite or any other application as long as source code is available.

59 BASIC BIOLOGICAL SCIENCES↗
NASA NTRS1967

A conceptual, parallel operating data compression processor

Data compressor processor concept envisions a simplified system for telemetry communications. It is simultaneously a zero-order processor and a floating aperture, a variable aperture, and a binary integer aperture with a decoded buffet fullness counter.

Anderson, T. O.↗
NASA NTRS1995

Parallel Climate Data Assimilation PSAS Package Achieves 18 GFLOPs on 512-Node Intel Paragon

Several algorithms were added to the Physical-space Statistical Analysis System (PSAS) from Goddard, which assimilates observational weather data by correcting for different levels of uncertainty about the data and different locations for mobile observation platforms. The new algorithms and use of the 512-node Intel Paragon allowed a hundred-fold decrease in processing time.

weather prediction climate modeling data assimilat↗
NASA NTRS2018

Systems and Methods for Transmitting Data Using Parallel Decode Capacity Optimized Symbol Constellations

Transmitters are described that use non-uniformly spaced symbol constellations that have increased capacity compared to conventional constellations. In many embodiments, a transmitter includes a coder configured to receive bits and output encoded bits, a mapper configured to map said encoded bits to symbols in a non-uniform symbol constellation selected from a plurality of symbol constellations, and a modulator configured to generate a signal for transmission via the communication channel using symbols generated by the mapper. In a variety of embodiments, wherein the non-uniform symbol constellation comprises a set of non-uniformly spaced constellation points, and the location points in the non-uniform symbol constellation are chosen to optimize parallel decode capacity of the non-uniform symbol constellation subject to at least one optimization constraint. In many embodiments, the non-uniform symbol constellation is a quadrature amplitude modulation constellation or a phase shift keyed constellation.

Barsoum, Maged F.↗
DOE OSTI2021

Improving scalability of parallel CNN training by adaptively adjusting parameter update frequency

Synchronous SGD with data parallelism, the most popular parallelization strategy for CNN training, suffers from the expensive communication cost of averaging gradients among all workers. The iterative parameter updates of SGD cause frequent communications and it becomes the performance bottleneck. In this paper, we propose a lazy parameter update algorithm that adaptively adjusts the parameter update frequency to address the expensive communication cost issue. Our algorithm accumulates the gradients if the difference of the accumulated gradients and the latest gradients is sufficiently small. Here, the less frequent parameter updates reduce the per-iteration communication cost while maintaining the model accuracy. Our experimental results demonstrate that the lazy update method remarkably improves the scalability while maintaining the model accuracy. For ResNet50 training on ImageNet, the proposed algorithm achieves a significantly higher speedup (739.6 on 2048 Cori KNL nodes) as compared to the vanilla synchronous SGD (276.6) while the model accuracy is almost not affected (<0.2% difference).

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

Scalable training of graph convolutional neural networks for fast and accurate predictions of HOMO-LUMO gap in molecules

Abstract Graph Convolutional Neural Network (GCNN) is a popular class of deep learning (DL) models in material science to predict material properties from the graph representation of molecular structures. Training an accurate and comprehensive GCNN surrogate for molecular design requires large-scale graph datasets and is usually a time-consuming process. Recent advances in GPUs and distributed computing open a path to reduce the computational cost for GCNN training effectively. However, efficient utilization of high performance computing (HPC) resources for training requires simultaneously optimizing large-scale data management and scalable stochastic batched optimization techniques. In this work, we focus on building GCNN models on HPC systems to predict material properties of millions of molecules. We use HydraGNN, our in-house library for large-scale GCNN training, leveraging distributed data parallelism in PyTorch. We use ADIOS, a high-performance data management framework for efficient storage and reading of large molecular graph data. We perform parallel training on two open-source large-scale graph datasets to build a GCNN predictor for an important quantum property known as the HOMO-LUMO gap. We measure the scalability, accuracy, and convergence of our approach on two DOE supercomputers: the Summit supercomputer at the Oak Ridge Leadership Computing Facility (OLCF) and the Perlmutter system at the National Energy Research Scientific Computing Center (NERSC). We present our experimental results with HydraGNN showing (i) reduction of data loading time up to 4.2 times compared with a conventional method and (ii) linear scaling performance for training up to 1024 GPUs on both Summit and Perlmutter.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

PandAna: A Python Analysis Framework for Scalable High Performance Computing in High Energy Physics

Modern experiments in high energy physics analyze millions of events recorded in particle detectors to select the events of interest and make measurements of physics parameters. These data can often be stored as tabular data in files with detector information and reconstructed quantities. Current techniques for event selection in these files lack the scalability needed for high performance computing environments. We describe our work to develop a high energy physics analysis framework suitable for high performance computing. This new framework utilizes modern tools for reading files and implicit data parallelism. Framework users analyze tabular data using standard, easy-to-use data analysis techniques in Python while the framework handles the file manipulations and parallelism without the user needing advanced experience in parallel programming. In future versions, we hope to provide a framework that can be utilized on a personal computer or a high performance computing cluster with little change to the user code.

Groh, Micah↗
NASA NTRS1993

The alignment-distribution graph

Implementing a data-parallel language such as Fortran 90 on a distributed-memory parallel computer requires distributing aggregate data objects (such as arrays) among the memory modules attached to the processors. The mapping of objects to the machine determines the amount of residual communication needed to bring operands of parallel operations into alignment with each other. We present a program representation called the alignment distribution graph that makes these communication requirements explicit. We describe the details of the representation, show how to model communication cost in this framework, and outline several algorithms for determining object mappings that approximately minimize residual communication.

Chatterjee, Siddhartha↗
NASA NTRS1993

The alignment-distribution graph

Implementing a data-parallel language such as Fortran 90 on a distributed-memory parallel computer requires distributing aggregate data objects (such as arrays) among the memory modules attached to the processors. The mapping of objects to the machine determines the amount of residual communication needed to bring operands of parallel operations into alignment with each other. We present a program representation called the alignment-distribution graph that makes these communication requirements explicit. We describe the details of the representation, show how to model communication cost in this framework, and outline several algorithms for determining object mappings that approximately minimize residual communication.

Chatterjee, Siddhartha↗
NASA NTRS1974

Data processor with conditionally supplied clock signals

Parallel data processor clock pulses are conditionally supplied to processing unit in response to relative values of binary bit of control source and binary bit derived on single lead. Use of single lead simplifies fabrication of large-scale integrated networks.

Lesniewski, R. J.↗
NASA NTRS1991

An efficient massively parallel Euler solver for unstructured grids

A data parallel mesh-vertex upwind finite-volume scheme for solving the Euler equations on triangular unstructured meshes is described. A novel vertex-based partitioning of the problem is introduced which minimizes the computation and communication costs associated with distributing the computation to the processors of a massively parallel computer. Finally, the performance of this unstructured computation on 8K processors of the Connection Machine CM-2 is compared with one processor of a Cray-YMP. The experiments show that 8K processors of the CM-2 achieve approximately 70 percent of the performance of one processor of the Cray-YMP on the unstructured mesh computations described here.

Hammond, Steven W.↗