Search NASA⌕ Search

SEARCH · Search NASA

Results for “Speedup”

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 55 records · Page 3
All sourcesNASA NTRSDOE OSTI
DOE OSTI2010

PEBBLES Mechanics Simulation Speedup

Pebble bed reactors contain large numbers of spherical fuel elements arranged randomly. Determining the motion and location of these fuel elements is required for calculating certain parameters of pebble bed reactor operation. These simulations involve hundreds of thousands of pebbles and involve determining the entire core motion as pebbles are recirculated. Single processor algorithms for this are insufficient since they would take decades to centuries of wall-clock time. This paper describes the process of parallelizing and speeding up the PEBBLES pebble mechanics simulation code. Both shared memory programming with the Open Multi-Processing API and distributed memory programming with the Message Passing Interface API are used in simultaneously in this process. A new shared memory lock-less linear time collision detection algorithm is described. This method allows faster detection of pebbles in contact than generic methods. These combine to make full recirculations on AVR sized reactors possible in months of wall clock time.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2013

Extending molecular simulation time scales: Parallel in time integrations for high-level quantum chemistry and complex force representations

Parallel in time simulation algorithms are presented and applied to conventional molecular dynamics (MD) and ab initio molecular dynamics (AIMD) models of realistic complexity. Assuming that a forward time integrator, f (e.g., Verlet algorithm), is available to propagate the system from time t{sub i} (trajectory positions and velocities x{sub i} = (r{sub i}, v{sub i})) to time t{sub i+1} (x{sub i+1}) by x{sub i+1} = f{sub i}(x{sub i}), the dynamics problem spanning an interval from t{sub 0}…t{sub M} can be transformed into a root finding problem, F(X) = [x{sub i} − f(x{sub (i−1})]{sub i} {sub =1,M} = 0, for the trajectory variables. The root finding problem is solved using a variety of root finding techniques, including quasi-Newton and preconditioned quasi-Newton schemes that are all unconditionally convergent. The algorithms are parallelized by assigning a processor to each time-step entry in the columns of F(X). The relation of this approach to other recently proposed parallel in time methods is discussed, and the effectiveness of various approaches to solving the root finding problem is tested. We demonstrate that more efficient dynamical models based on simplified interactions or coarsening time-steps provide preconditioners for the root finding problem. However, for MD and AIMD simulations, such preconditioners are not required to obtain reasonable convergence and their cost must be considered in the performance of the algorithm. The parallel in time algorithms developed are tested by applying them to MD and AIMD simulations of size and complexity similar to those encountered in present day applications. These include a 1000 Si atom MD simulation using Stillinger-Weber potentials, and a HCl + 4H{sub 2}O AIMD simulation at the MP2 level. The maximum speedup ((serial execution time)/(parallel execution time) ) obtained by parallelizing the Stillinger-Weber MD simulation was nearly 3.0. For the AIMD MP2 simulations, the algorithms achieved speedups of up to 14.3. The parallel in time algorithms can be implemented in a distributed computing environment using very slow transmission control protocol/Internet protocol networks. Scripts written in Python that make calls to a precompiled quantum chemistry package (NWChem) are demonstrated to provide an actual speedup of 8.2 for a 2.5 ps AIMD simulation of HCl + 4H{sub 2}O at the MP2/6-31G* level. Implemented in this way these algorithms can be used for long time high-level AIMD simulations at a modest cost using machines connected by very slow networks such as WiFi, or in different time zones connected by the Internet. The algorithms can also be used with programs that are already parallel. Using these algorithms, we are able to reduce the cost of a MP2/6-311++G(2d,2p) simulation that had reached its maximum possible speedup in the parallelization of the electronic structure calculation from 32 s/time step to 6.9 s/time step.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2016

Reaching bandwidth saturation using transparent injection parallelization

Although logically available, applications may not exploit enough instantaneous communication concurrency to maximize network utilization on HPC systems. This is exacerbated in hybrid programming models that combine single program multiple data with OpenMP or CUDA. We present the design of a “multi-threaded” runtime able to transparently increase the instantaneous network concurrency and to provide near saturation bandwidth, independent of the application configuration and dynamic behavior. The runtime offloads communication requests from application level tasks to multiple communication servers. The servers use system specific performance models to attain network saturation. Our techniques alleviate the need for spatial and temporal application level message concurrency optimizations. Experimental results show improved message throughput and bandwidth by as much as 150% for 4 KB messages on InfiniBand and by as much as 120% for 4 KB messages on Cray Aries. For more complex operations such as all-to-all collectives, we observe as much as 30% speedup. This translates into 23% speedup on 12,288 cores for a NAS FT implemented using FFTW. We observe as much as 76% speedup on 1500 cores for an already optimized UPC+OpenMP geometric multigrid application using hybrid parallelism. For the geometric multigrid GPU implementation, we observe as much as 44% speedup on 512 GPUs.

Chaimov, Nicholas↗
DOE OSTI2011

DOE SBIR Phase-1 Report on Hybrid CPU-GPU Parallel Development of the Eulerian-Lagrangian Barracuda Multiphase Program

This report gives the result from the Phase-1 work on demonstrating greater than 10x speedup of the Barracuda computer program using parallel methods and GPU processors (General-Purpose Graphics Processing Unit or Graphics Processing Unit). Phase-1 demonstrated a 12x speedup on a typical Barracuda function using the GPU processor. The problem test case used about 5 million particles and 250,000 Eulerian grid cells. The relative speedup, compared to a single CPU, increases with increased number of particles giving greater than 12x speedup. Phase-1 work provided a path for reformatting data structure modifications to give good parallel performance while keeping a friendly environment for new physics development and code maintenance. The implementation of data structure changes will be in Phase-2. Phase-1 laid the ground work for the complete parallelization of Barracuda in Phase-2, with the caveat that implemented computer practices for parallel programming done in Phase-1 gives immediate speedup in the current Barracuda serial running code. The Phase-1 tasks were completed successfully laying the frame work for Phase-2. The detailed results of Phase-1 are within this document. In general, the speedup of one function would be expected to be higher than the speedup of the entire code because of I/O functions and communication between the algorithms. However, because one of the most difficult Barracuda algorithms was parallelized in Phase-1 and because advanced parallelization methods and proposed parallelization optimization techniques identified in Phase-1 will be used in Phase-2, an overall Barracuda code speedup (relative to a single CPU) is expected to be greater than 10x. This means that a job which takes 30 days to complete will be done in 3 days. Tasks completed in Phase-1 are: Task 1: Profile the entire Barracuda code and select which subroutines are to be parallelized (See Section Choosing a Function to Accelerate) Task 2: Select a GPU consultant company and jointly parallelize subroutines (CPFD chose the small business EMPhotonics for the Phase-1 the technical partner. See Section Technical Objective and Approach) Task 3: Integrate parallel subroutines into Barracuda (See Section Results from Phase-1 and its subsections) Task 4: Testing, refinement, and optimization of parallel methodology (See Section Results from Phase-1 and Section Result Comparison Program) Task 5: Integrate Phase-1 parallel subroutines into Barracuda and release (See Section Results from Phase-1 and its subsections) Task 6: Roadmap of Phase-2 (See Section Plan for Phase-2) With the completion of Phase 1 we have the base understanding to completely parallelize Barracuda. An overview of the work to move Barracuda to a parallelized code is given in Plan for Phase-2.

02 PETROLEUM↗
DOE OSTI2022

VWC-BERT: Scaling Vulnerability–Weakness–Exploit Mapping on Modern AI Accelerators

Defending cybersystems needs accurate mapping of software and hardware vulnerabilities to generalized descriptions of weaknesses, and weaknesses to exploits. These mappings enable cyber defenders to build plans for effective defense and assessment of potential risks to a cybersystem. With close to 170k vulnerabilities, manual mapping is not a feasible option. However, automated mapping is challenging due to limited training data, computational intractability, and limitations in computational natural language processing. Tools based on breakthroughs in Transformer-based language models have been demonstrated to classify vulnerabilities with high accuracy. We make three key contributions in this paper: (1) We present a new framework, \VWCBERT, that augments the Transformer-based hierarchical multi-class classification framework of Das et al. (\textsc{V2W-BERT}) with the ability to map weaknesses to exploits. (2) We implement \VWCBERT~ on modern AI accelerator platforms using two data parallel techniques for the pre-training phase and demonstrate nearly linear speedups across NVIDIA and Graphcore accelerator platforms. We observe nearly linear speedups for up to 16 V100 and 8 A100 GPUs, and about 3.4$\times$ speedup for A100 relative to V100 GPUs. We also observe excellent speedups on Graphcore, with $5.7\times$ speedup on 128 IPUs relative to 16 IPUs. Enabled by scaling, we also demonstrate higher accuracy using a larger language model, RoBERTa-Large. We show up to 87\% accuracy for strict and up to 98\% accuracy for relaxed classification. (3) We develop a novel parallel link manager for the link prediction phase and demonstrate up to 21$\times$ speedup with 16 V100 GPUs relative to one V100 GPU, and thus reducing the runtime from 2.5 hours to 10 minutes. We believe that generalizability and scalability of \VWCBERT~ will benefit both the theoretical development and practical deployment of novel cyberdefense solutions and vulnerability classification.

Das, Siddhartha Shankar↗
DOE OSTI2016

Accelerating the Mining of Influential Nodes in Complex Networks through Community Detection

Computing the set of influential nodes with a given size to ensure maximal spread of influence on a complex network is a challenging problem impacting multiple applications. A rigorous approach to influence maximization involves utilization of optimization routines that comes with a high computational cost. In this work, we propose to exploit the existence of communities in complex networks to accelerate the mining of influential seeds. We provide intuitive reasoning to explain why our approach should be able to provide speedups without significantly degrading the extent of the spread of influence when compared to the case of influence maximization without using the community information. Additionally, we have parallelized the complete workflow by leveraging an existing parallel implementation of the Louvain community detection algorithm. We then conduct a series of experiments on a dataset with three representative graphs to first verify our implementation and then demonstrate the speedups. Our method achieves speedups ranging from 3x - 28x for graphs with small number of communities while nearly matching or even exceeding the activation performance on the entire graph. Complexity analysis reveals that dramatic speedups are possible for larger graphs that contain a correspondingly larger number of communities. In addition to the speedups obtained from the utilization of the community structure, scalability results show up to 6.3x speedup on 20 cores relative to the baseline run on 2 cores. Finally, current limitations of the approach are outlined along with the planned next steps.

community detection↗
DOE OSTI2012

Large-scale Nanostructure Simulations from X-ray Scattering Data On Graphics Processor Clusters

X-ray scattering is a valuable tool for measuring the structural properties of materialsused in the design and fabrication of energy-relevant nanodevices (e.g., photovoltaic, energy storage, battery, fuel, and carbon capture andsequestration devices) that are key to the reduction of carbon emissions. Although today's ultra-fast X-ray scattering detectors can provide tremendousinformation on the structural properties of materials, a primary challenge remains in the analyses of the resulting data. We are developing novelhigh-performance computing algorithms, codes, and software tools for the analyses of X-ray scattering data. In this paper we describe two such HPCalgorithm advances. Firstly, we have implemented a flexible and highly efficient Grazing Incidence Small Angle Scattering (GISAXS) simulation code based on theDistorted Wave Born Approximation (DWBA) theory with C++/CUDA/MPI on a cluster of GPUs. Our code can compute the scattered light intensity from any givensample in all directions of space; thus allowing full construction of the GISAXS pattern. Preliminary tests on a single GPU show speedups over 125x compared tothe sequential code, and almost linear speedup when executing across a GPU cluster with 42 nodes, resulting in an additional 40x speedup compared to usingone GPU node. Secondly, for the structural fitting problems in inverse modeling, we have implemented a Reverse Monte Carlo simulation algorithm with C++/CUDAusing one GPU. Since there are large numbers of parameters for fitting in the in X-ray scattering simulation model, the earlier single CPU code required weeks ofruntime. Deploying the AccelerEyes Jacket/Matlab wrapper to use GPU gave around 100x speedup over the pure CPU code. Our further C++/CUDA optimization deliveredan additional 9x speedup.

36 MATERIALS SCIENCE↗
DOE OSTI2016

Space-Dependent Wielandt Shift Methods for Multigroup Diffusion Eigenvalue Problems

In the calculation of neutronic eigenfunctions and eigenvalues, it is often desirable to approximate the neutron transport equation by a diffusion equation that requires significantly less computational resources to solve. Even when a full transport solution is needed, the solution of the diffusion system can be used to accelerate iterative methods for calculating the transport eigenfunction and eigenvalue. The convergence rate of the basic power iteration method for eigenvalue problems is governed by the dominance ratio; realistic reactor core problems often have high dominance ratios and the power iteration method can converge slowly as a result. A common technique for accelerating the convergence of power iteration is the Wielandt shift (WS) method, which reduces the dominance ratio by 'shifting' the eigenvalue spectrum using an estimate of the desired eigenvalue. WS provides a significant speedup over power iteration if the eigenvalue estimate is close to the true eigenvalue, but it can be difficult to obtain such an estimate. In practice, the shift in the WS method can be dynamically updated in the iteration scheme, in a way that depends on the eigenvalue estimates from the most recent iterations. We present a physically-motivated Space-Dependent Wielandt Shift (SDWS) for solving multigroup diffusion eigenvalue problems. The space-dependence of the eigenvalue shift in SDWS enables a more effective acceleration of the power iteration scheme than WS at the beginning of the iteration scheme, when the current estimate of the eigenvalue is not close to the actual eigenvalue. Several variants of SDWS are discussed here, including an SDWS method that combines the strengths of SDWS (in the early iterations) with those of standard WS (in the later iterations). Numerical results from a 1-D multigroup diffusion problem demonstrate an order of magnitude speedup over power iteration and a 41% speedup over a standard WS technique. This speedup is problem-dependent and appears to be greater for 'more difficult' problems. In this summary, the term 'speedup' refers to a reduction in the number of iterations required to converge to the correct eigenvalue and eigenfunction. Because the scattering is treated fully implicitly (i.e., it is part of the left-hand side of the diffusion system and directly inverted), the computational cost of a single iteration does not vary significantly among the methods discussed in this summary. The number of iterations is therefore representative of the total computational cost for a method. We acknowledge that certain iterative solvers for linear systems (e.g., GMRES) require more computational time to converge when the system is nearly singular (i.e., when a very aggressive Wielandt shift is applied). Nonetheless, when using SDWS in the manner described in this summary, we have seen that the reduction in computational time from reducing the number of power iterations generally outweighs the extra computational time required by iterative solvers such as GMRES. (authors)

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2017

eQE: An open‐source density functional embedding theory code for the condensed phase

Abstract In this work, we present the main features and algorithmic details of a novel implementation of the frozen density embedding (FDE) formulation of subsystem density functional theory (DFT) that is specifically designed to enable ab initio molecular dynamics (AIMD) simulations of large‐scale condensed‐phase systems containing 1000s of atoms. This code (available at http://eqe.rutgers.edu ) has been given the moniker of embedded Quantum ESPRESSO (eQE) as it is a generalization of the open‐source Quantum ESPRESSO (QE) suite of programs. The strengths of eQE reside in a hierarchical parallelization scheme that allows for an efficient and fully self‐consistent treatment of the electronic structure (via the addition of an additional DIIS extrapolation layer) while simultaneously exploiting the inherent symmetries and periodicities in the system (via sampling of subsystem‐specific first Brillouin zones and utilization of subsystem‐specific basis sets). While bulk liquids and molecular crystals are two classes of systems that exemplify the utility of the FDE approach (as these systems can be partitioned into weakly interacting subunits), we show that eQE has significantly extended this regime of applicability by outperforming standard semilocal Kohn–Sham DFT (KS‐DFT) for large‐scale heterogeneous catalysts with quite different layer‐specific electronic structure and intrinsic periodicities. eQE features very favorable strong parallel scaling for a model system of bulk liquid water composed of 256 water molecules, which allows for a significant decrease in the overall time to solution when compared to KS‐DFT. We show that eQE achieves speedups greater than one order of magnitude ( ) when performing AIMD simulations of such large‐scale condensed‐phase systems as: (1) molecular liquids via bulk liquid water represented by 1024 independent water molecules (3072 atoms with a 25.3× speedup over KS‐DFT), (2) polypeptide/biomolecule solvation via ( gly ) 6 solvated in (H 2 O) 395 (1230 atoms with a 38.6× speedup over KS‐DFT), and (3) molecular crystals via a 3 × 3 × 3 periodic supercell of pentacene (1940 atoms with a 12.0× speedup over KS‐DFT). These results represent a significant improvement over the current state‐of‐the‐art and now enable subsystem DFT‐based AIMD simulations of realistically sized condensed‐phase systems of interest throughout chemistry, physics, and materials science.

Genova, Alessandro↗
DOE OSTI2020

A space-time parallel algorithm with adaptive mesh refinement for computational fluid dynamics

This work describes a space-time parallel algorithm with space-time adaptive mesh refinement (AMR). AMR with subcycling is added to multigrid reduction-in-time (MGRIT) in order to provide solution efficient adaptive grids with a reduction in work performed on coarser grids. This algorithm is achieved by integrating two software libraries: XBraid (Parallel time integration with multigrid. https://computation.llnl.gov/projects/parallel-timeintegration-multigrid) and Chombo (Chombo software package for AMR applications—design document, 2014). The former is a parallel time integration library using multigrid and the latter is a massively parallel structured AMR library. Employing this adaptive space-time parallel algorithm is Chord (Comput Fluids 123:202–217, 2015), a computational fluid dynamics (CFD) application code for solving compressible fluid dynamics problems. For the same solution accuracy, speedups are demonstrated from the use of space-time parallelization over the time-sequential integration on Couette flow and Stokes’ second problem. On a transient Couette flow case, at least a 1.5× speedup is achieved, and with a time periodic problem, a speedup of up to 13.7× over the time-sequential case is obtained. In both cases, the speedup is achieved by adding processors and exploring additional parallelization in time. The numerical experiments show the algorithm is promising for CFD applications that can take advantage of the time parallelism. Future work will focus on improving the parallel performance and providing more tests with complex fluid dynamics to demonstrate the full potential of the algorithm.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

A machine learning approach for increased throughput of density functional theory substitutional alloy studies

In this study, a machine learning-based technique is developed to reduce the computational cost required to explore large design spaces of substitutional alloys. The first advancement is based on a neural network (NN) approach to predict the initial position of minority and majority ions prior to DFT relaxation. The second advancement is to allow the NN to predict the total energy for every possibility minority ion position and select the most stable configuration in the absence of relaxing each trial minority configuration. A bismuth oxide materials system, (Bi x La y Yb z ) 2 MoO 6 , was used as the model system to demonstrate the developed methods and quantify the resulting computational speedup. Compared to a brute force method that requires the calculation of every permutation of minority configuration and subsequent DFT relaxation, a 1.3× speedup was realized if the NN predicted the initial configuration of ions prior to relaxation. Implementation of the second advancement allowed the NN to predict the total energy for all possible trial configurations and downselect the most stable configurations prior to relaxation, resulting in a speedup of approximately 37×. Validation was done by comparing position and energy between the NN and DFT predictions. A maximum position vector mean squared error (MSE) of 1.6 × 10 -2 and a maximum energy MSE of 2.3 × 10 -7 was predicted for the worst case configuration. This method demonstrates a significant computational speedup, which has the potential for even greater computational savings for larger compositional design spaces.

36 MATERIALS SCIENCE↗
DOE OSTI2011

Modeling of 10 GeV-1 TeV laser-plasma accelerators using Lorentz boosted simulations

Modeling of laser-plasma wakefield accelerators in an optimal frame of reference [J.-L. Vay, Phys. Rev. Lett. 98, 130405 (2007)] allows direct and efficient full-scale modeling of deeply depleted and beam loaded laser-plasma stages of 10 GeV-1 TeV (parameters not computationally accessible otherwise). This verifies the scaling of plasma accelerators to very high energies and accurately models the laser evolution and the accelerated electron beam transverse dynamics and energy spread. Over 4, 5, and 6 orders of magnitude speedup is achieved for the modeling of 10 GeV, 100 GeV, and 1 TeV class stages, respectively. Agreement at the percentage level is demonstrated between simulations using different frames of reference for a 0.1 GeV class stage. Obtaining these speedups and levels of accuracy was permitted by solutions for handling data input (in particular, particle and laser beams injection) and output in a relativistically boosted frame of reference, as well as mitigation of a high-frequency instability that otherwise limits effectiveness.

43 PARTICLE ACCELERATORS↗
DOE OSTI2019

Multiply accelerated ReaxFF molecular dynamics: coupling parallel replica dynamics with collective variable hyper dynamics

To address the time scales required to study complex chemical reactions, methods performing accelerated molecular dynamics are necessary even with the recent advancement in high-performance computing. A number of different acceleration techniques are available. Here we explore potential synergies between two popular acceleration methods – Parallel Replica Dynamics (PRD) and Collective Variable Hyperdynamics (CVHD), by analysing the speedup obtained for the pyrolysis of n-dodecane. We determine that PRD + CVHD provides additional speedup to CVHD by reaching the required time scales for the reaction at an earlier wall-clock time. Although some speedup is obtained with the additional replicas, we found that the effectiveness of the inclusion of PRD is depreciated for systems where there is a dramatic increase in reaction rates induced by CVHD. Similar observations were made in the simulation of ethylene-carbonate/Li system, which is inherently more reactive than pyrolysis, suggest that the speedup obtained via the combination of the two acceleration methods can be generalised to most practical chemical systems.

42 ENGINEERING↗
DOE OSTI2020

Wind Farm Simulation and Layout Optimization in Complex Terrain: Preprint

This work reports on incorporating complex terrain into wind farm simulations for the purpose of layout optimization. Adding complex terrain boundary conditions to NREL's medium fidelity computational fluid dynamics model, WindSE, produces significant separation, flow curvature, and speedup effects that would otherwise be difficult to capture with lower-fidelity models or a flat-terrain assumption. These flow features, in turn, can significantly impact the optimal turbine array layout. We demonstrate the impact of complex terrain on flow in both an idealized and real-world setting, and discuss modifications to the code that enable gradient-based optimization using terrain-aware adjoint gradients. Through several optimization case studies, we show that the layout optimization process takes advantage of speedup effects on terrain high points, and leverages flow curvature effects that modify wake trajectories. This yields substantial power improvements over gridded layouts, and hints at future research directions in simulation and optimization for wake trajectories in complex terrain.

17 WIND ENERGY↗
DOE OSTI2019

Fine-Grained Exploitation of Mixed Precision for Faster CNN Training

As deep convolutional neural networks (CNNs) have become increasingly popular and successful at an ever-widening number of machine learning tasks specialized hardware has become increasingly available for training and deploying them. NVIDIA's recent Volta architecture includes tensor cores which perform a fused operation reduced and mixed precision (16-bit multiply, 32-bit accumulate). Recent research indicates that, typically, very little is lost (in terms of training accuracy) when half precision is used in place of single precision, and performance gains can be made by doing arithmetic in reduced precision. In this work we demonstrate that making layer-by-layer choices as to the arithmetic/data precision can lead to further performance improvement. In our study of 25,200 CNNs we demonstrate an average speedup (over purely half precision) of 1.27x and speedups as high as 3.64x by appropriately combining single and half precision arithmetic and data types on a layer-by-layer basis.

Johnston, Travis↗
DOE OSTI2022

Automatic Generation of High-Performance Convolution Kernels on ARM CPUs for Deep Learning

In this work, we present FastConv, a template-based code auto-generation open source library that can automatically generate high-performance deep learning convolution kernels of arbitrary matrices/tensors shapes. FastConv is based on the Winograd algorithm, which is reportedly the highest performing algorithm for the time-consuming convolution layers of convolutional neural networks. ARM CPUs cover a wide range designs and specifications, from embedded devices to HPC-grade CPUs. The leads to the dilemma of how to consistently optimize Winograd-based convolution solvers for convolution layers of different shapes. FastConv addresses this problem by using templates to auto-generate multiple shapes of tuned kernels variants suitable for skinny tall matrices. As a performance portable library, FastConv transparently searches for the best combination of kernel shapes, cache tiles, scheduling of loop orders, packing strategies, access patterns, and online/offline computations. Auto-tuning is used to search the parameter configuration space for the best performance for a given target architecture and problem size. The experiments with layer-wise evaluation on the VGG--16 model confirms a 1.25x performance gains is got by tuning the Winograd library. Integrated comparison results shows 1.02x to 1.40x, 1.14x to 2.17x, and 1.22x and 2.48x speedup is achieved over NNPACK, Arm NN, and FeatherCNN on the Kunpeng 920 beside few cases. Furthermore, problem size performance portability experiments with various convolution shapes shows that FastConv achieves 1.2x to 1.7x speedup and 2x to 22x speedup over NNPACK and ARM NN inference engine using Winograd on Kunpeng 920 . CPU performance portability evaluation on the VGG--16 show an average speedup over NNPACK of 1.42x, 1.21x, 1.26x, 1.37x, 2.26x, and 11.02x is observed on Kunpeng 920, Snapdragon 835, 855, 888, Apple M1, and AWS Graviton2, respectively.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

The Minos Computing Library: Efficient Parallel Programming for Extremely Heterogeneous Systems

Hardware specialization has become the silver bullet to achieve efficient high performance, from Systems-on-Chip systems, where hardware specialization can be ``extreme'', to large-scale HPC systems. As the complexity of the systems increases, so does the complexity of programming such architectures in a portable way. This work introduces the Minos Computing Library (MCL), as system software, programming model, and programming model runtime that facilitate programming extremely heterogeneous systems. MCL supports the execution of several multi-threaded applications within the same compute node, performs asynchronous execution of application tasks, efficiently balances computation across hardware resources, and provides performance portability. We show that code developed on a personal desktop automatically scales up to fully utilize powerful workstations with 8 GPUs and down to power-efficient embedded systems. MCL provides up to 17.5x speedup over OpenCL on NVIDIA DGX-1 systems and up to 1.88x speedup on single-GPU systems. In multi-application workloads, MCL dynamically resource allocation provides up to 2.43x performance improvement over manual, static allocation of computing resources.

Gioiosa, Roberto↗
DOE OSTI2021

Advances in Mixed Precision Algorithms: 2021 Edition

Over the last year, the ECP xSDK-multiprecision effort has made tremendous progress in developing and deploying new mixed precision technology and customizing the algorithms for the hardware deployed in the ECP flagship supercomputers. The effort also has succeeded in creating a cross-laboratory community of scientists interested in mixed precision technology and now working together in deploying this technology for ECP applications. In this report, we highlight some of the most promising and impactful achievements of the last year. Among the highlights we present are: Mixed precision IR using a dense LU factorization and achieving a 1.8× speedup on Spock; results and strategies for mixed precision IR using a sparse LU factorization; a mixed precision eigenvalue solver; Mixed Precision GMRES-IR being deployed in Trilinos, and achieving a speedup of 1.4× over standard GMRES; compressed Basis (CB) GMRES being deployed in Ginkgo and achieving an average 1.4× speedup over standard GMRES; preparing hypre for mixed precision execution; mixed precision sparse approximate inverse preconditioners achieving an average speedup of 1.2×; and detailed description of the memory accessor separating the arithmetic precision from the memory precision, and enabling memory-bound low precision BLAS 1/2 operations to increase the accuracy by using high precision in the computations without degrading the performance. We emphasize that many of the highlights presented here have also been submitted to peer-reviewed journals or established conferences, and are under peer-review or have already been published.

97 MATHEMATICS AND COMPUTING↗