Tuning parallel data compression and I/O for large-scale earthquake simulation
Explore the source record for details and available documents.
Mathematics and Computing, Computer science↗
SEARCH · Search NASA
Results for “parallel, scaling”
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.
Spectral quadrature for the first principles study of crystal defects: Application to magnesium
In this work, we present an accurate and efficient finite-difference formulation and parallel implementation of Kohn-Sham Density (Operator) Functional Theory (DFT) for non periodic systems embedded in a bulk environment. Specifically, employing non-local pseudopotentials, local reformulation of electrostatics, and truncation of the spatial Kohn-Sham Hamiltonian, and the Linear Scaling Spectral Quadrature method to solve for the pointwise electronic fields in real-space and the non-local component of the atomic force, we develop a parallel finite difference framework suitable for distributed memory computing architectures to simulate non-periodic systems embedded in a bulk environment. Choosing examples from magnesium-aluminum alloys, we first demonstrate the convergence of energies and forces with respect to spectral quadrature polynomial order, and the width of the spatially truncated Hamiltonian. Next, we demonstrate the parallel scaling of our framework, and show that the computation time and memory scale linearly with respect to the number of atoms. Next, we use the developed framework to simulate isolated point defects and their interactions in magnesium-aluminum alloys. Our findings conclude that the binding energies of divacancies, Al solute-vacancy and two Al solute atoms are anisotropic and are dependent on cell size. Furthermore, the binding is favorable in all three cases.
71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
Towards Native Execution of Deep Learning on a Leadership-Class HPC System
Large parallel machines generally offer the best parallel performance with "native execution" that is achieved using codes developed with the optimized compilers, communication libraries, and runtimes offered on the machines. In this paper, we report and analyze performance results from native execution of deep learning on a leadership-class high-performance computing (HPC) system. Using our new code called DeepEx, we present a study of the parallel speed up and convergence rates of learning achieved with native parallel execution. In the trade-off between computational parallelism and synchronized convergence, we first focus on maximizing parallelism while still obtaining convergence. Scaling results are reported from execution on up to 15,000 GPUs using two scientific data sets from atom microscopy and protein folding applications, and also using the popular ImageNet data set. In terms of the traditional measure of parallel speed up, excellent scaling is observed up to 12,000 GPUs. Additionally, accounting for convergence rates of deep learning accuracy or error, a deep learning-specific metric called "learning speed up" is also tracked. The performance results indicate the need to evaluate parallel deep learning execution in terms of learning speed up, and point to additional directions for improved exploitation of high-end HPC systems.
Yoginath, Srikanth↗
libEnsemble: A complete Python toolkit for dynamic ensembles of calculations
Almost all science and engineering applications eventually stop scaling: their runtime no longer decreases as available computational resources increase. Therefore, many applications will struggle to efficiently use emerging extreme-scale high-performance, parallel, and distributed systems. libEnsemble is a complete Python toolkit and workflow system for intelligently driving ensembles of experiments or simulations at massive scales. It enables and encourages multidisciplinary design, decision, and inference studies portably running on laptops, clusters, and supercomputers.
97 MATHEMATICS AND COMPUTING↗
Integrated Model, Batch and Domain Parallelism in Training Neural Networks
We propose a new integrated method of exploiting model, batch and domain parallelism for the training of deep neural networks (DNNs) on large distributed-memory computers using minibatch stochastic gradient descent (SGD). Our goal is to find an efficient parallelization strategy for a fixed batch size using $P$ processes. Our method is inspired by the communication-avoiding algorithms in numerical linear algebra. We see $P$ processes as logically divided into a $P_r \times P_c$ grid where the $P_r$ dimension is implicitly responsible for model/domain parallelism and the $P_c$ dimension is implicitly responsible for batch parallelism. In practice, the integrated matrix-based parallel algorithm encapsulates these types of parallelism automatically. We analyze the communication complexity and analytically demonstrate that the lowest communication costs are often achieved neither with pure model nor with pure data parallelism. We also show how the domain parallel approach can help in extending the theoretical scaling limit of the typical batch parallel method.
97 MATHEMATICS AND COMPUTING↗
DFT-FE 1.0: A massively parallel hybrid CPU-GPU density functional theory code using finite-element discretization
In this work, we present DFT-FE 1.0, building on DFT-FE 0.6 [Comput. Phys. Commun. 246, 106853 (2020)], to conduct fast and accurate large-scale density functional theory (DFT) calculations (reaching ~ 100,000 electrons) on both many-core CPU and hybrid CPU-GPU computing architectures. This work involves improvements in the real-space formulation—via an improved treatment of the electrostatic interactions that substantially enhances the computational efficiency—as well high-performance computing aspects, including the GPU acceleration of all the key compute kernels in DFT-FE. We demonstrate the accuracy by comparing the ground-state energies, ionic forces and cell stresses on a wide-range of benchmark systems against those obtained from widely used DFT codes. Further, we demonstrate the numerical efficiency of our implementation, which yields ~ 20× CPU-GPU speed-up by using GPU acceleration on hybrid CPU-GPU nodes. Notably, owing to the parallel-scaling of the GPU implementation, we obtain wall-times of 80–140 seconds for full ground-state calculations, with stringent accuracy, on benchmark systems containing ~ 6, 000 – 15,000 electrons.
pseudopotential↗
Extreme-scale stochastic optimization and simulation via learning-enhanced decomposition and parallelization (Final Technical Report)
Stochastic optimization and simulation models ubiquitously arise in designing and operating complex service/engineering systems. They can be extreme in scale due to high-dimensional data and decisions, and can also involve decisions made sequentially in response to newly revealed data, both causing significant computational challenge. The objective of this research is to explore a unified framework that integrates machine learning with discrete optimization and risk-averse modeling, to improve the efficiency of decomposition paradigms for stochastic optimization and simulations at extreme scale. The models we consider represent a broad class of complex decision-making problems, where 0-1 or continuous decisions are made before and/or after knowing multiple sources of uncertainties that could be correlated. We will employ machine learning methods to dynamically decide and prioritize computational procedures, including cut generation, branching, and bounding of the optimal objective. Furthermore, the research will shed new lights on the traditional decomposition algorithms for extreme-scale computing. Deliverables of the research include new modeling and computational methods for advancing the state-of-the-art research in optimization and simulation, bringing many relevant risk-averse, data-driven optimization problems in practice within the range of tractability. Examples include distributed computing server scheduling and sensor deployment for monitoring critical infrastructures. Success in this effort will enable progress in solving multiple extreme-scale problems in the complex system design and operations arising from DoE missions in energy, environment, and national security.
24 POWER TRANSMISSION AND DISTRIBUTION↗
Aspherical particle models for molecular dynamics simulation
In traditional molecular dynamics (MD) simulations, atoms and coarse-grained particles are modeled as point masses interacting via isotropic potentials. For studies where particle shape plays a vital role, more complex models are required. Here, we describe a spectrum of approaches for modeling aspherical particles, all of which are now available (some recently) as options within the LAMMPS MD package. Broadly these include two classes of models. In the first, individual particles are aspherical, either via a pairwise anisotropic potential which implicitly assigns a simple geometric shape to each particle, or in a more general way where particles store internal state which can explicitly define a complex geometric shape. In the second class of models, individual particles are simple points or spheres, but rigid body constraints are used to create composite aspherical particles in a variety of complex shapes. This work discusses parallel algorithms and associated data structures for both kinds of models, which enable dynamics simulations of aspherical particle systems across a wide range of length and time scales. We also highlight parallel performance and scalability and give a few illustrative examples of aspherical models in different contexts.
74 ATOMIC AND MOLECULAR PHYSICS↗
Parallel multiple-chain DRAM MCMC for large-scale geosteering inversion and uncertainty quantification
Geosteering is the proactive control of a wellbore placement based on the downhole measurements, aiming at maximizing economic production from the well. We report the recent development of azimuthal resistivity logging-whiledrilling (LWD) tools have a much larger depth of detection, thus sets higher demands for geosteering inversion as more unknown parameters need to be taken into account in the earth model to be inverted. For complicated nonlinear problems, traditional deterministic inversion methods are more likely to be trapped near local optima. In general, Markov chain Monte Carlo (MCMC) inversion methods are more capable of finding the global optimal solution and providing additional uncertainty analyses by sampling from the target distribution. However, MCMC methods usually have the problem of slow convergence. Even though optimization methods like delayed rejection (DR) and adaptive Metropolis (AM) were proposed to speed up the convergence, using MCMC methods to solve geosteering inversion problems may still incur unacceptable time cost. To reduce the sampling cost of MCMC methods, we use parallel multiple-chain DRAM MCMC methods to solve geosteering inverse problems and estimate the corresponding uncertainty. A clustering method, density-based spatial clustering of applications with noise (DBSCAN), is applied to select the best solution among many results. Lastly, the simulation results show that running many relatively short MCMC chains for one problem can obtain a similar result as running a long single chain. Besides, avoiding communications between multiple Markov chains during the sampling process yields almost linear scalability.
97 MATHEMATICS AND COMPUTING↗
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↗
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↗
Ion temperature and rotation fluctuation measurements with ultra-fast charge exchange recombination spectroscopy (UF-CHERS) in the DIII-D tokamak
An upgraded detector and several optimizations have significantly improved the Ultra-Fast Charge Exchange Recombination Spectroscopy (UF-CHERS) diagnostic sensitivity to ion temperature and parallel velocity fluctuations at turbulence relevant spatio-temporal scales. Normalized broadband ion temperature and parallel velocity fluctuations down to x̃x∼1% (x = Ti, v∥) and up to ∼450 kHz have been measured in a variety of plasmas. The multi-field nature of the CHERS technique also allows measurements of the cross-phase angles of the fluctuating fields. UF-CHERS is optimized to observe emissions from the electron exchange reaction between intrinsic C6+ and hydrogenic neutral beam injected particles near 529 nm. UF-CHERS consists of two chords separated by ∼1 cm radially, less than the turbulence correlation length in DIII-D plasmas, which enables correlated measurements to suppress incoherent electronic and photon noise. The optical components of the spectrometer include a volume-phase-holographic grating with >90% transmission between 528 and 530 nm and f/2 200-mm lenses, selected to maximize the optical efficiency and photon flux. Diffracted light from each chord is collected in eight spectral bins, each with a bandwidth of ∼0.25 nm, and detected and amplified by chilled avalanche photodiodes and custom high-gain, wide bandwidth low-noise preamplifiers to achieve the optimal signal-to-noise ratio. The resulting signals are digitized at 1 MHz, 103–104× faster than the conventional CHERS diagnostics. Spatial coverage is achieved by repositioning a motorized fiber tray between plasmas. UF-CHERS measurements will advance the understanding of turbulent ion transport and contribute to the validation of transport models and simulations.
Truong, D. D. (ORCID:0000000285732539)↗
On the use of a multigrid-reduction-in-time algorithm for multiscale convergence of turbulence simulations
Simulations of turbulent flow present challenges in terms of accuracy and affordability on modern highly-parallel computer architectures. A multigrid-reduction-in-time algorithm is used to provide a framework for separately evolving different scales of turbulence and for parallelizing the temporal domain, thereby increasing the concurrency. It is hypothesized that the space–time locality of the small scales of turbulence can be used to circumvent difficulties in applying temporal multigrid to flows dominated by inertial physics. For algorithms that fall well short of spectral accuracy (fourth-order is used in this work) attention must be paid to the accuracy of features on scales transferred between multigrid levels. Numerical experiments were performed using implicit large-eddy simulation. Results from applying the approach to an infinite-Reynolds number Taylor–Green flow and a double-shear flow at a Reynolds number of 11650 provide strong evidence that the approach has merit. The multigrid-reduction-in-time framework can be used to parallelize the temporal domain of a high-Reynolds-number turbulent flow and permit independent convergence of different scales. Establishing this foundation allows for future research in reducing the wall-clock time to solve turbulent flows while retaining the same accuracy as sequential solvers. In conclusion, current performance results from parallelizing the temporal domain are not competitive with those from sequential-in-time methods.
97 MATHEMATICS AND COMPUTING↗
TriC: Distributed-memory Triangle Counting by Exploiting the Graph Structure
Graph analytics has emerged as an important tool in the analysis of large scale data from diverse application domains such as social networks, cyber security and bioinformatics. Counting the number of triangles in a graph is a fundamental kernel with several applications such as detecting the community structure of a graph or in identifying important vertices in a graph. The ubiquity of massive datasets is driving the need to scale graph analytics on parallel systems. However, numerous challenges exist in efficiently parallelizing graph algorithms, especially on distributed-memory systems. Irregular memory accesses and communication patterns, low computation to communication ratios, and the need for frequent synchronization are some of the leading challenges. In this paper, we present TriC, our distributed-memory implementation of triangle counting in graphs using the Message Passing Interface (MPI), as a submission to the 2020 GraphChallenge competition. Using a set of synthetic and real-world inputs from the challenge, we demonstrate a speedup of up to 90x relative to previous work on 32 processor-cores of a NERSC Cori node. We also provide details from distributed runs with up to8192 processes along with strong scaling results. The observations presented in this work provide an understanding of the system-level bottlenecks at scale that specifically impact sparse-irregular workloads and will therefore benefit other efforts to parallelize graph algorithms.
Halappanavar, Mahantesh↗
Efficient quadrature rules for finite element discretizations of nonlocal equations
In this paper we design efficient quadrature rules for finite element discretizations of nonlocal diffusion problems with compactly supported kernel functions. Two of the main challenges in nonlocal modeling and simulations are the prohibitive computational cost and the nontrivial implementation of discretization schemes, especially in three-dimensional settings. In this work we circumvent both challenges by introducing a parametrized mollifying function that improves the regularity of the integrand, utilizing an adaptive integration technique, and exploiting parallelization. We first showthat the “mollified” solution converges to the exact one as the mollifying parameter vanishes, then we illustrate the consistency and accuracy of the proposed method on several two- and three-dimensional test cases. Furthermore, we demonstrate the good scaling properties of the parallel implementation of the adaptive algorithm and we compare the proposed method with recently developed techniques for efficient finite element assembly.
97 MATHEMATICS AND COMPUTING↗
Efficient quadrature rules for finite element discretizations of nonlocal equations
Here, in this paper, we design efficient quadrature rules for finite element (FE) discretizations of nonlocal diffusion problems with compactly supported kernel functions. Two of the main challenges in nonlocal modeling and simulations are the prohibitive computational cost and the nontrivial implementation of discretization schemes, especially in three-dimensional settings. In this work, we circumvent both challenges by introducing a parametrized mollifying function that improves the regularity of the integrand, utilizing an adaptive integration technique, and exploiting parallelization. We first show that the “mollified” solution converges to the exact one as the mollifying parameter vanishes, then we illustrate the consistency and accuracy of the proposed method on several two- and three-dimensional test cases. Furthermore, we demonstrate the good scaling properties of the parallel implementation of the adaptive algorithm and we compare the proposed method with recently developed techniques for efficient FE assembly.
97 MATHEMATICS AND COMPUTING↗
HipMCL: a high-performance parallel implementation of the Markov clustering algorithm for large-scale networks
Biological networks capture structural or functional properties of relevant entities such as molecules, proteins or genes. Characteristic examples are gene expression networks or protein–protein interaction networks, which hold information about functional affinities or structural similarities. Such networks have been expanding in size due to increasing scale and abundance of biological data. While various clustering algorithms have been proposed to find highly connected regions, Markov Clustering (MCL) has been one of the most successful approaches to cluster sequence similarity or expression networks. Despite its popularity, MCL’s scalability to cluster large datasets still remains a bottleneck due to high running times and memory demands. In this paper, we present High-performance MCL (HipMCL), a parallel implementation of the original MCL algorithm that can run on distributed-memory computers. We show that HipMCL can efficiently utilize 2000 compute nodes and cluster a network of ~70 million nodes with ~68 billion edges in ~2.4 h. By exploiting distributed-memory environments, HipMCL clusters large-scale networks several orders of magnitude faster than MCL and enables clustering of even bigger networks. Finally, HipMCL is based on MPI and OpenMP and is freely available under a modified BSD license.
59 BASIC BIOLOGICAL SCIENCES↗
Highly scalable distributed-memory sparse triangular solution algorithms.
This paper presents a highly efficient distributed-memory parallel sparse triangular solver. The triangular solution phase is often performed following factorization phase in the sparse linear solvers and has become increasingly computationally expensive for direct solvers with many right hand sides (RHSs) or preconditioned iterative solvers. However, the low arithmetic intensity and sequential nature of the triangular solve algorithm pose performance challenges for its large-scale distributed-memory parallelization. In this work, we propose several strategies to enhance scalability of an algorithm with 2D block cyclic process layout. First, an asynchronous binary-tree-based communication scheme implemented via non-blocking MPI functions is leveraged to broadcast partial solutions and reduce partial updates among a subset of processes for each block column and row of the triangular matrix, respectively. This scheme reduces message latency, improves communication load balance and significantly accelerates asynchronous execution of the triangular solve. In addition, efficient BLAS operations and threading implementations are exploited to accelerate local computations and further reduce process idle time. The proposed strategies are implemented in SuperLU_DIST and numerical experiments show up to 4.4x improvement with one right-hand side and up to 6.1x improvement with 50 right- hand sides on 4096 processes, compared to the current release. This is the first time that sparse triangular solution is demonstrated strong scaling on more than 4000 cores.
Liu, Yang↗