Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “high performance parallel computing”

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 901 records · Page 50
All sourcesNASA NTRSDOE OSTI
DOE OSTI2014

WE-G-16A-01: Evolution of Radiation Treatment Planning

Welcome and Introduction - Lawrence N. Rothenberg This symposium is one a continuing series of presentations at AAPM Annual Meetings on the historical aspects of medical physics, radiology, and radiation oncology that have been organized by the AAPM History Committee. Information on previous presentations including “Early Developments in Teletherapy” (Indianapolis 2013), “Historical Aspects of Cross-Sectional Imaging” (Charlotte 2012), “Historical Aspects of Brachytherapy” (Vancouver 2011), “50 Years of Women in Medical Physics” (Houston 2008), and “Roentgen's Early Investigations” (Minneapolis 2007) can be found in the Education Section of the AAPM Website. The Austin 2014 History Symposium will be on “Evolution of Radiation Treatment Planning.” Overview - Radhe Mohan Treatment planning is one of the most critical components in the chain of radiation therapy of cancers. Treatment plans of today contain a wide variety of sophisticated information conveying the potential clinical effectiveness of the designed treatment to practitioners. Examples of such information include dose distributions superimposed on three- or even four-dimensional anatomic images; dose volume histograms, dose, dose-volume and dose-response indices for anatomic structures of interest; etc. These data are used for evaluating treatment plans and for making treatment decisions. The current state-of-the-art has evolved from the 1940s era when the dose to the tumor and normal tissues was estimated approximately by manual means. However, the symposium will cover the history of the field from the late-1950's, when computers were first introduced for treatment planning, to the present state involving the use of high performance computing and advanced multi-dimensional anatomic, functional and biological imaging, focusing only on external beam treatment planning. The symposium will start with a general overview of the treatment planning process including imaging, structure delineation, assignment of dose requirements, consideration of uncertainties, selection of beam configurations and shaping of beams, and calculations, optimization and evaluation of dose distributions. This will be followed by three presentations covering the evolution of treatment planning, which parallels the evolution of computers, availability of advanced volumetric imaging and the development of novel technologies such as dynamic multi-leaf collimators and online image guidance. This evolution will be divided over three distinct periods - prior to 1970's, the 2D era; from 1980 to the mid-1990's, the 3D era; and from the mid 1990's to today, the IMRT era. When the World was Flat: The Two-Dimensional Radiation Therapy Era” - Jacob Van Dyk In the 2D era, anatomy was defined with the aid of solder wires, special contouring devices and projection x-rays. Dose distributions were calculated manually from single field, flat surface isodoses on transparencies. Precalculated atlases of generic dose distributions were produced by the International Atomic Energy Agency. Massive time-shared main frames and mini-computers were used to compute doses at individual points or dose distributions in a single plane. Beam shapes were generally rectangular, with wedges, missing tissue compensators and occasional blocks to shield critical structures. Dose calculations were measurement-based or they used primary and scatter calculations based on scatter-air ratio methodologies. Dose distributions were displayed on line printers as alpha-numeric character maps or isodose patterns made with pen plotters. More than Pretty Pictures: 3D Treatment Planning and Conformal Therapy - Benedick A. Fraass The introduction of computed tomography allowed the delineation of anatomy three-dimensionally and, supported partly by contracts from the National Cancer Institute, made possible the introduction and clinical use of 3D treatment planning, leading to development and use of 3D conformal therapy in the 1980's. 3D computer graphics and 3D anatomical structure definitions made possible Beam's Eye View (BEV) displays, making conformal beam shaping and much more sophisticated beam arrangements possible. These conformal plans significantly improved target dose coverage as well as normal tissue sparing. The use of dose volume histograms, gross/clinical/planning target volumes, MRI and PET imaging, multileaf collimators, and computer-controlled treatment delivery made sophisticated planning approaches practical. The significant improvements in dose distributions and analysis achievable with 3D conformal therapy made possible formal dose escalation and normal tissue tolerance clinical studies that set new and improved expectations for improved local control and decreasing complications in many clinical sites. From the Art to the State of the Art: Inverse Planning and IMRT - Thomas R. Bortfeld While the potential of intensity modulation was recognized in the mid- 1980's, intensity-modulated radiotherapy (IMRT) did not become a reality until the mid-1990's. Broad beams of photons could be sub-divided into narrow beamlets whose intensities could be determined using sophisticated optimization algorithms to appropriately balance tumor dose with normal tissue sparing. The development of dynamic multi-leaf collimators (on conventional linear accelerators as well as in helical delivery devices) enabled the efficient delivery of IMRT. The evolution of IMRT planning is continuing in the form of Volumetric Modulated Arc Therapy (VMAT) and through advanced optimization tools, such as multi-criteria optimization, automated IMRT planning, and robust optimization to protect dose distributions against uncertainties. IMRT also facilitates “dose painting” in which different sub-volumes of the target are prescribed different doses. Clearly, these advancements are being made possible by the increasing power and lower cost of computers and developments in other fields such as imaging and operations research. Summary - Radhe Mohan The history does not end here. The advancement of treatment planning is expected to continue, leading to further automation and improvements in conformality and robustness of dose distributions, particularly in the area of particle therapy. Radiobiological modeling will gain emphasis as part of the planning process. Learning Objectives: The scope of changes in technology and the capabilities of radiation treatment planning The impact of these changes in the quality of treatment plans and optimality of dose distributions The impact of development in other fields (imaging, computers, operations research, etc.) on the evolution of radiation treatment planning.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2022

Spinor $GW$ Bethe-Salpeter calculations in BerkeleyGW: Implementation, symmetries, benchmarking, and performance

Computing the GW quasiparticle band structure and Bethe-Salpeter equation (BSE) absorption spectra for materials with spin-orbit coupling have commonly been done by treating GW corrections and spin-orbit coupling (SOC) as separate perturbations to density-functional theory. However, accurate treatment of materials with strong spin-orbit coupling (such as many topological materials of recent interest, and thermoelectrics) often requires a nonperturbative approach using spinor wave functions in the Kohn-Sham equation and GW/BSE. Such calculations have only recently become available, in particular for the BSE. Here, we have implemented this approach in the plane-wave pseudopotential GW/BSE code BerkeleyGW, which is highly parallelized and widely used in the electronic-structure community. We present reference results for quasiparticle band structures and optical absorption spectra of solids with different strengths of spin-orbit coupling, including Si, Ge, GaAs, GaSb, CdSe, Au, and Bi 2 Se 3 . The calculated quasiparticle band gaps of these systems are found to agree with experiment to within a few tens of meV. SOC splittings are found to be generally in better agreement with experiment, including quasiparticle corrections to band energies. The absorption spectrum of GaAs is not significantly impacted by the inclusion of spin-orbit coupling due to its relatively small value (0.2 eV) in the Λ direction, while the absorption spectrum of GaSb calculated with the spinor GW/BSE captures the large spin-orbit splitting of peaks in the spectrum. For the prototypical topological insulator Bi 2 Se 3 , we find a drastic change in the low-energy band structure compared to that of DFT, with the spinorial treatment of the GW approximation correctly capturing the parabolic nature of the valence and conduction bands after including off-diagonal self-energy matrix elements. We present the detailed methodology, approach to spatial symmetries for spinors, comparison against other codes, and performance compared to spinless GW/BSE calculations and perturbative approaches to SOC. This work aims to spur further development of spinor GW/BSE methodology in excited-state research software and enables a more accurate and detailed exploration of electronic and optical properties of materials containing elements with large atomic numbers.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2017

The Pointing Self-calibration Algorithm for Aperture Synthesis Radio Telescopes

This paper is concerned with algorithms for calibration of direction-dependent effects (DDE) in aperture synthesis radio telescopes (ASRT). After correction of direction-independent effects (DIE) using self-calibration, imaging performance can be limited by the imprecise knowledge of the forward gain of the elements in the array. In general, the forward gain pattern is directionally dependent and varies with time due to a number of reasons. Some factors, such as rotation of the primary beam with Parallactic Angle for Azimuth–Elevation mount antennas are known a priori. Some, such as antenna pointing errors and structural deformation/projection effects for aperture-array elements cannot be measured a priori. Thus, in addition to algorithms to correct for DD effects known a priori, algorithms to solve for DD gains are required for high dynamic range imaging. Here, we discuss a mathematical framework for antenna-based DDE calibration algorithms and show that this framework leads to computationally efficient optimal algorithms that scale well in a parallel computing environment. As an example of an antenna-based DD calibration algorithm, we demonstrate the Pointing SelfCal (PSC) algorithm to solve for the antenna pointing errors. Our analysis show that the sensitivity of modern ASRT is sufficient to solve for antenna pointing errors and other DD effects. We also discuss the use of the PSC algorithm in real-time calibration systems and extensions for antenna Shape SelfCal algorithm for real-time tracking and corrections for pointing offsets and changes in antenna shape.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2021

Towards performance portability in the Spark astrophysical magnetohydrodynamics solver in the Flash-X simulation framework

Simulations of core-collapse supernovae, and other astrophysical phenomena, are quintessential extreme-scale computing challenges. For core-collapse supernova simulations to be carried out by the ExaStar project under the Exascale Computing Project umbrella, a robust, efficient, and state-of-the-art magnetohydrodynamics solver is a critical requirement. In Flash-X, the primary software instrument for ExaStar, a new magnetohydrodynamics solver has been designed and implemented from the ground up to achieve accuracy and efficiency for simulations of complex astrophysical flows. This new solver, dubbed Spark, uses high-order spatial reconstruction, Runge-Kutta time integration, and an efficient cell-centered approach to satisfying the divergence-free condition for the magnetic fields. Spark was written to be optimized for data locality in cache hierarchy of CPUs. Since data locality optimizations for cache hierarchy are not directly compatible with those of accelerators, we have taken the approach of using program synthesis to avoid massive amounts of code replication that would be necessary if we were to maintain two different versions of the solver. Our program synthesis relies on a simple key-dictionary approach, implemented in python, that enables us to assemble the version of the solver suitable for the target hardware from code fragments identified by specific keys. In this work, we describe the data locality optimizations of the solver for CPUs and accelerators and the program synthesis tools that enable this portability. We also detail the parallel performance of Spark for both CPUs and accelerators.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2014

A Screen Space GPGPU Surface LIC Algorithm for Distributed Memory Data Parallel Sort Last Rendering Infrastructures

The surface line integral convolution(LIC) visualization technique produces dense visualization of vector fields on arbitrary surfaces. We present a screen space surface LIC algorithm for use in distributed memory data parallel sort last rendering infrastructures. The motivations for our work are to support analysis of datasets that are too large to fit in the main memory of a single computer and compatibility with prevalent parallel scientific visualization tools such as ParaView and VisIt. By working in screen space using OpenGL we can leverage the computational power of GPUs when they are available and run without them when they are not. We address efficiency and performance issues that arise from the transformation of data from physical to screen space by selecting an alternate screen space domain decomposition. We analyze the algorithm's scaling behavior with and without GPUs on two high performance computing systems using data from turbulent plasma simulations.

Loring, Burlen↗
DOE OSTI2023

Advanced Algal Biofoundries for the Production of Polyurethane Precursors

The primary goal of the BEEPs project was to develop a process that could accelerate the development of algae as bioproduction platforms, from initial chemical product concept to an economically viable market supply. Under this program we elected to develop strains of algae that could generate polyurethane precursors, while simultaneously developing basic genetic tools to enable improved algal production systems. This program was specifically designed to incorporate National Laboratories as a means to utilize the expertise and facilities for new bio-production platforms. To that end, we designed a program to collaborate with the Agile BioFoundry at Lawrence Berkeley National Laboratory (LBNL) and computational platforms at Pacific Northwest National Laboratory (PNNL). In addition to these National Laboratory partners, we also had academic partners from UC Davis and Georgia Tech, as well as commercial partners Algenesis Materials and BASF. To achieve these goals, we initially focused on developing the genetic tools and high throughput screening technologies necessary to generate and assess production of polymer precursors (succinic acid) in algae and cyanobacteria, including advanced promoters and biosensors. In parallel, we computationally identified potential production bottlenecks and then used the developed genetic tools to increase production rates and yields. Constant feedback of data was used in conjunction with machine learning, high-throughput cell sorting, and synthetic biology, for additional targeted metabolic engineering. Multiple rounds of tool design, building, testing, and learning were supplied to partners at PNNL and LBNL to develop new models and tools that could expedite bioproduction platform development and increase yield performance. We targeted, and achieved, the FOA requirement yield metric of 20 g/L, as a milestone and deliverable from at least one of our engineered strains for the production of succinic acid.

09 BIOMASS FUELS↗
DOE OSTI2016

An Asynchronous Task-based Fan-Both Sparse Cholesky Solver

Systems of linear equations arise at the heart of many scientific and engineering applications. Many of these linear systems are sparse; i.e., most of the elements in the coefficient matrix are zero. Direct methods based on matrix factorizations are sometimes needed to ensure accurate solutions. For example, accurate solution of sparse linear systems is needed in shift-invert Lanczos to compute interior eigenvalues. The performance and resource usage of sparse matrix factorizations are critical to time-to-solution and maximum problem size solvable on a given platform. In many applications, the coefficient matrices are symmetric, and exploiting symmetry will reduce both the amount of work and storage cost required for factorization. When the factorization is performed on large-scale distributed memory platforms, communication cost is critical to the performance of the algorithm. At the same time, network topologies have become increasingly complex, so that modern platforms exhibit a high level of performance variability. This makes scheduling of computations an intricate and performance-critical task. In this paper, we investigate the use of an asynchronous task paradigm, one-sided communication and dynamic scheduling in implementing sparse Cholesky factorization (symPACK) on large-scale distributed memory platforms. Our solver symPACK relies on efficient and flexible communication primitives provided by the UPC++ library. Performance evaluation shows good scalability and that symPACK outperforms state-of-the-art parallel distributed memory factorization packages, validating our approach on practical cases.

Jacquelin, Mathias↗
DOE OSTI2022

ExaCA: A performance portable exascale cellular automata application for alloy solidification modeling

Modeling the as-solidified grain structures that form during alloy processing is a critical component in understanding process-property relationships, particularly for additive manufacturing (AM) where grain structure is very sensitive to processing conditions. While cellular automata (CA)-based models have proven able to predict aspects of microstructure for several alloys and AM process conditions, long run times and large resource sets required limit the utility and the problem size to which existing CA models can be applied. As part of the ExaAM project, an initiative within the Exascale Computing Project (ECP) to develop, test, and optimize an exascale-capable coupled and self-consistent model of AM parts, we developed ExaCA (https://github.com/LLNL/ExaCA) for the liquid–solid phase transformation in the wake of AM melt pools. The CA-based code is parallelized using MPI and the Kokkos programming model, the latter enabling simulation on both CPUs and GPUs within a single-source implementation. Here, we detail the steps taken to transform a baseline, MPI-based CA code into one that is performant on CPUs and GPUs. Performance testing of ExaCA on Summit (a pre-exascale machine at Oak Ridge National Laboratory) was used to quantify CPU–GPU speedup comparing with equal numbers of nodes. Testing showed comparable CPU performance to the MPI-only CA code and a 5-20x speedup when running AM-based test problems using GPUs. The improved performance of CA through GPU utilization and the performance portable nature of ExaCA will enable accurate part-scale modeling by harnessing the power of current and future generations of high performance computing resources. Future work will include improving the strong scaling of ExaCA on GPUs by reducing load imbalance associated with the locality of the problem, and continuing performance optimization across exascale hardware.

36 MATERIALS SCIENCE↗
DOE OSTI2023

Energetic particle marginal stability profile for HL-2M integrated simulation based on neural network module

Abstract A critical gradient model is employed to develop a module of energetic particle (EP) marginal stability profiles in OMFIT integrated simulations for studying EP transport. Currently, each iteration of transport evolution is approximately 10 min in the integrated simulation, whereas, the EP marginal stability profile, which serves as an input in the integrated simulation could take much longer; the reason being a combination of the TGLFEP and EPtran codes is employed in our previous investigation. To reduce the simulation time, the critical gradient is predicted by a neural network instead of the TGLFEP code, and the EPtran code is revised with parallel computing, so that the running time of this module can be controlled to within 5 min. The predictions are in good agreement with previous approaches. The integrated simulation of HL-2M with Alfven eigenmodes transported by neutral beam EP profiles indicates that EP transport reduces the total pressure and current as expected, but could also under some conditions raise the safety factor in the core, which is favorable for reversed magnetic shear and high-performance plasmas.

Physics↗
DOE OSTI2016

The Reverse Cuthill-McKee Algorithm in Distributed-Memory

Ordering vertices of a graph is key to minimize fill-in and data structure size in sparse direct solvers, maximize locality in iterative solvers, and improve performance in graph algorithms. Except for naturally parallelizable ordering methods such as nested dissection, many important ordering methods have not been efficiently mapped to distributed-memory architectures. In this paper, we present the first-ever distributed-memory implementation of the reverse Cuthill-McKee (RCM) algorithm for reducing the profile of a sparse matrix. Our parallelization uses a two-dimensional sparse matrix decomposition. We achieve high performance by decomposing the problem into a small number of primitives and utilizing optimized implementations of these primitives. Our implementation shows strong scaling up to 1024 cores for smaller matrices and up to 4096 cores for larger matrices.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2019

Exploring MPI Communication Models for Graph Applications Using Graph Matching as a Case Study

Traditional implementations of parallel graph operations on distributed memory platforms are written using Message Passing Interface (MPI) point to point communication primitives such as Send-Recv (blocking and nonblocking). Apart from this classical model, the MPI community has over the years added other communication models; however, their suitability for handling the irregular traffic workloads typical of graph operations remain comparatively less explored. Our aim in this paper is to study these relatively underutilized communication models of MPI for graph applications. More specifically, we evaluate MPI’s one-sided programming, or Remote Memory Access (RMA), and nearest neighborhood collectives using a process graph topology. There are features in these newer models that are intended to better map to irregular communication patterns, as exemplified in graph algorithms. As a concrete application for our case study, we use distributed memory implementations of a weighted graph matching algorithm to investigate performances of MPI-3 RMA and neighborhood collective operations compared to nonblocking Send- Recv. A matching in a graph is a subset of edges such that no two matched edges are incident on the same vertex. A maximum weight matching is a matching of maximum weight computed as the sum of the weights of matched edges. Algorithms for optimal matching have quadratic complexity and are difficult to parallelize. More importantly, execution of graph matching is dominated by high volume of irregular memory accesses, making it an ideal candidate for studying the effects of various MPI communication models on graph applications at scale. Our neighborhood collectives and RMA implementations yields up to 6× speedup over traditional optimized nonblocking Send-Recv implementations, on thousands of cores of the NERSC Cori supercomputer. We believe the lessons learned from this case study can be adopted to benefit a wider range of graph application codes.

Ghosh, Sayan↗
DOE OSTI2017

Social energy: mining energy from the society

The inherent nature of energy, i.e., physicality, sociality and informatization, implies the inevitable and intensive interaction between energy systems and social systems. From this perspective, we define 'social energy' as a complex sociotechnical system of energy systems, social systems and the derived artificial virtual systems which characterize the intense intersystem and intra-system interactions. The recent advancement in intelligent technology, including artificial intelligence and machine learning technologies, sensing and communication in Internet of Things technologies, and massive high performance computing and extreme-scale data analytics technologies, enables the possibility of substantial advancement in socio-technical system optimization, scheduling, control and management. In this paper, we provide a discussion on the nature of energy, and then propose the concept and intention of social energy systems for electrical power. A general methodology of establishing and investigating social energy is proposed, which is based on the ACP approach, i.e., 'artificial systems' (A), 'computational experiments' (C) and 'parallel execution' (P), and parallel system methodology. A case study on the University of Denver (DU) campus grid is provided and studied to demonstrate the social energy concept. In the concluding remarks, we discuss the technical pathway, in both social and nature sciences, to social energy, and our vision on its future.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2022

High-dimensional discrete Fourier transform gates with a quantum frequency processor

The discrete Fourier transform (DFT) is of fundamental interest in photonic quantum information, yet the ability to scale it to high dimensions depends heavily on the physical encoding, with practical recipes lacking in emerging platforms such as frequency bins. In this article, we show that d -point frequency-bin DFTs can be realized with a fixed three-component quantum frequency processor (QFP), simply by adding to the electro-optic modulation signals one radio-frequency harmonic per each incremental increase in d . We verify gate fidelity F W > 0.9997 and success probability P W > 0.965 up to d = 10 in numerical simulations, and experimentally implement the solution for d = 3, utilizing measurements with parallel DFTs to quantify entanglement and perform tomography of multiple two-photon frequency-bin states. Our results furnish new opportunities for high-dimensional frequency-bin protocols in quantum communications and networking.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Xyce™ Parallel Electronic Simulator Users’ Guide (V.7.7)

This manual describes the use of the Xyce Parallel Electronic Simulator. Xyce has been designed as a SPICE-compatible, high-performance analog circuit simulator, and has been written to support the simulation needs of the Sandia National Laboratories electrical designers.

42 ENGINEERING↗
DOE OSTI2024

A Full-Stack Exploration of Language-Based Parallelism in Fortran 2023

This poster explores native parallel features in Fortran 2023 through the lens of supporting applications with libraries, compilers, and parallel runtimes. The language revision informally named Fortran 2008 introduced parallelism in the form of Single Program Multiple Data (SPMD) execution with two broad feature sets: (1) loop-level parallelism via do concurrent and (2) a Partitioned Global Address Space (PGAS) comprised of distributed “coarray” data structures. Fortran’s native parallelism has demonstrated high performance [1] and reduced the burden of inserting what sometimes amounts to more directives than code. Several compilers support both feature sets, typically by translating do concurrent into serial do loops annotated by parallel directives and by translating SPMD/PGAS features into direct calls to a communication library. Our research focuses primarily on two questions: (1) can the compiler’s parallel runtime library be developed in the language being compiled (Fortran) and (2) can we define an interface to the runtime that liberates compilers from being hardwired to one runtime and vice versa. We are answering these questions by developing the Parallel Runtime Interface for Fortran (PRIF) [2] and the Co-Array Fortran Framework of Efficient Interfaces to Network Environments (Caffeine) [3]. Caffeine is initially targeting adoption by LLVM Flang, a new open-source Fortran compiler developed by a broad community in industry, academia, and government labs. We are also exploring the use of these features in Inference-Engine, a deep learning library designed to facilitate neural network training and inference for high-performance computing applications written in modern Fortran.

Rasmussen, Katherine↗
DOE OSTI2014

Magnetic shielding of Hall thrusters at high discharge voltages

A series of numerical simulations and experiments have been performed to assess the effectiveness of magnetic shielding in a Hall thruster operating in the discharge voltage range of 300–700 V (I{sub sp} ≈ 2000–2700 s) at 6 kW, and 800 V (I{sub sp} ≈ 3000) at 9 kW. At 6 kW, the magnetic field topology with which highly effective magnetic shielding was previously demonstrated at 300 V has been retained for all other discharge voltages; only the magnitude of the field has been changed to achieve optimum thruster performance. It is found that magnetic shielding remains highly effective for all discharge voltages studied. This is because the channel is long enough to allow hot electrons near the channel exit to cool significantly upon reaching the anode. Thus, despite the rise of the maximum electron temperature in the channel with discharge voltage, the electrons along the grazing lines of force remain cold enough to eliminate or reduce significantly parallel gradients of the plasma potential near the walls. Computed maximum erosion rates in the range of 300–700 V are found not to exceed 10{sup −2} mm/kh. Such rates are ∼3 orders of magnitude less than those observed in the unshielded version of the same thruster at 300 V. At 9 kW and 800 V, saturation of the magnetic circuit did not allow for precisely the same magnetic shielding topology as that employed during the 6-kW operation since this thruster was not designed to operate at this condition. Consequently, the maximum erosion rate at the inner wall is found to be ∼1 order of magnitude higher (∼10{sup −1} mm/kh) than that at 6 kW. At the outer wall, the ion energy is found to be below the sputtering yield threshold so no measurable erosion is expected.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2020

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↗
DOE OSTI2022

Distributed Training and Optimization of Neural Networks

Deep learning models are yielding increasingly better performances thanks to multiple factors. To be successful, model may have large number of parameters or complex architectures and be trained on large dataset. This leads to large requirements on computing resource and turn around time, even more so when hyperparameter optimization is done (e.g. search over model architectures). While this is a challenge that goes beyond particle physics, we review the various ways to do the necessary computations in parallel, and put it in the context of high-energy physics.

Vlimant, Jean-Roch↗