Search NASA⌕ Search

SEARCH · Search NASA

Results for “high performance and distributed 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 541 records · Page 30
All sourcesNASA NTRSDOE OSTI
DOE OSTI2012

Characterizing output bottlenecks in a supercomputer

Supercomputer I/O loads are often dominated by writes. HPC (High Performance Computing) file systems are designed to absorb these bursty outputs at high bandwidth through massive parallelism. However, the delivered write bandwidth often falls well below the peak. This paper characterizes the data absorption behavior of a center-wide shared Lustre parallel file system on the Jaguar supercomputer. We use a statistical methodology to address the challenges of accurately measuring a shared machine under production load and to obtain the distribution of bandwidth across samples of compute nodes, storage targets, and time intervals. We observe and quantify limitations from competing traffic, contention on storage servers and I/O routers, concurrency limitations in the client compute node operating systems, and the impact of variance (stragglers) on coupled output such as striping. We then examine the implications of our results for application performance and the design of I/O middleware systems on shared supercomputers.

Xie, Bing↗
DOE OSTI2025

Record acceleration of the two-dimensional Ising model using a high-performance wafer-scale engine

The versatility and wide-ranging applicability of the Ising model, originally introduced to study phase transitions in magnetic materials, have made it a cornerstone in statistical physics and a valuable tool for evaluating the performance of emerging computer hardware. Here, we present a novel implementation of the two-dimensional Ising model on Cerebras Wafer-Scale Engine (WSE) – a revolutionary processor that is opening new frontiers in computing. In our deployment of the checkerboard algorithm, we optimized the Ising model to take advantage of the unique WSE architecture. Specifically, we employed a compressed bit representation storing 16 spins on each int16 word, and efficiently distributed the spins over the processing units enabling seamless weak scaling and limiting communications to only immediate neighboring units. Our implementation can handle up to 754 simulations in parallel, achieving an aggregate of over 61.8 trillion flip attempts per second for Ising models with up to 200 million spins. This represents a gain of up to 148 times over previously reported single-devices with a highly optimized implementation on NVIDIA V100 and up to 88 times in productivity compared to NVIDIA H100. Our findings highlight the significant potential of the WSE in scientific computing, particularly in the field of materials modeling.

Ising model↗
DOE OSTI2014

A dosimetric evaluation of the Eclipse AAA algorithm and Millennium 120 MLC for cranial intensity-modulated radiosurgery

The aim of this study is to assess the accuracy of a convolution-based algorithm (anisotropic analytical algorithm [AAA]) implemented in the Eclipse planning system for intensity-modulated radiosurgery (IMRS) planning of small cranial targets by using a 5-mm leaf-width multileaf collimator (MLC). Overall, 24 patient-based IMRS plans for cranial lesions of variable size (0.3 to 15.1 cc) were planned (Eclipse, AAA, version 10.0.28) using fixed field-based IMRS produced by a Varian linear accelerator equipped with a 120 MLC (5-mm width on central leaves). Plan accuracy was evaluated according to phantom-based measurements performed with radiochromic film (EBT2, ISP, Wayne, NJ). Film 2D dose distributions were performed with the FilmQA Pro software (version 2011, Ashland, OH) by using the triple-channel dosimetry method. Comparison between computed and measured 2D dose distributions was performed using the gamma method (3%/1 mm). Performance of the MLC was checked by inspection of the DynaLog files created by the linear accelerator during the delivery of each dynamic field. The absolute difference between the calculated and measured isocenter doses for all the IMRS plans was 2.5% ± 2.1%. The gamma evaluation method resulted in high average passing rates of 98.9% ± 1.4% (red channel) and 98.9% ± 1.5% (blue and green channels). DynaLog file analysis revealed a maximum root mean square error of 0.46 mm. According to our results, we conclude that the Eclipse/AAA algorithm provides accurate cranial IMRS dose distributions that may be accurately delivered by a Varian linac equipped with a Millennium 120 MLC.

61 RADIATION PROTECTION AND DOSIMETRY↗
DOE OSTI2019

Exploratory Visual Analysis of Anomalous Runtime Behavior in Streaming High Performance Computing Applications

Online analysis of runtime behavior is essential for performance tuning in streaming scientific workflows. Integration of anomaly detection and visualization is necessary to support human-centered analysis, such as verification of candidate anomalies utilizing domain knowledge. In this work, we propose an efficient and scalable visual analytics system for online performance analysis of scientific workflows toward the exascale scenario. Here, our approach uses a call stack tree representation to encode the structural and temporal information of the function executions. Based on the call stack tree features (e.g., execution time of the root function or vector representation of the tree structure), we employ online anomaly detection approaches to identify candidate anomalous function executions. We also present a set of visualization tools for verification and exploration in a level-of-detailed manner. General information, such as distribution of execution times, are provided in an overview visualization. The detailed structure (e.g., function invocation relations) and the temporal information (e.g., message communication) of the execution call stack of interest are also visualized. The usability and efficiency of our methods are verified in a real-world HPC application.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Parallel algorithms for finding connected components using linear algebra

Finding connected components is one of the most widely used operations on a graph. Optimal serial algorithms for the problem have been known for half a century, and many competing parallel algorithms have been proposed over the last several decades under various different models of parallel computation. This paper presents a class of parallel connected-component algorithms designed using linear-algebraic primitives. These algorithms are based on a PRAM algorithm by Shiloach and Vishkin and can be designed using standard GraphBLAS operations. Here, we demonstrate two algorithms of this class, one named LACC for Linear Algebraic Connected Components, and the other named FastSV which can be regarded as LACC’s simplification. With the support of the highly-scalable Combinatorial BLAS library, LACC and FastSV outperform the previous state-of-the-art algorithm by a factor of up to 12x for small to medium scale graphs. For large graphs with more than 50B edges, LACC and FastSV scale to 4K nodes (262K cores) of a Cray XC40 supercomputer and outperform previous algorithms by a significant margin. This remarkable performance is accomplished by (1) exploiting sparsity that was not present in the original PRAM algorithm formulation, (2) using high-performance primitives of Combinatorial BLAS, and (3) identifying hot spots and optimizing them away by exploiting algorithmic insights.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Numerical assessment of triply periodic minimal surfaces for direct air capture of carbon dioxide

Direct air capture (DAC) systems often consist of packing material wetted by a capture fluid that reacts with CO 2 in the airstream. The efficiency of the contactor is determined by a complex relationship of fluid dynamics, heat and mass transfer, contactor geometry, and chemical properties. The efficiency of the contactor must be balanced with other factors, primarily pressure drop through the system. Triply periodic minimal surfaces (TPMS) are a class of differential surfaces that have been explored in multiple engineering applications and have been shown to exhibit excellent performance when used in heat exchangers. Their tortuous path provides a high surface-to-volume ratio and favorable trade-off between contact area and pressure drop. In this work, a gyroid-type TPMS contactor was evaluated using computational fluid dynamics for a variety of geometric parameters to explore the potential benefit of TPMS shapes for DAC applications. A thin-film model was employed to model the flow and distribution of the capture solvent, allowing efficient simulations of TPMS structures at scale by eliminating the need for a computationally intensive interface capturing method. A liquid-gas mass transfer model was implemented in the commercial software STAR-CCM+ and used to predict the CO 2 capture efficiency and study the trade-off between capture performance and pressure drop through analysis of capture rates, mass transfer coefficients, and other relevant variables. TPMS contactors with a variety of geometric parameters and two capture solvent options were investigated to determine the effect of design choices on the operational performance of DAC systems. In conclusion, results showed that while contactor geometry is the dominant factor in efficiency and pressure drop, the physiochemical properties of the solvent are an important secondary influence on the contactor performance.

CFD↗
DOE OSTI2024

Numerical simulation of involute-plate research reactor flow behavior using RANS, LES and DNS

This paper investigates the flow behavior of involute-plate research reactors by performing Reynolds-Averaged Navier Stokes simulation (RANS), Large Eddy Simulation (LES) and Direct Numerical Simulation (DNS) of the channel flow between fuel plates. By modeling turbulence with different numerical approaches, this study provides data with three levels of fidelity. For the RANS simulation, three widely used turbulence models, i.e., k-ε, k-ω, Reynolds Stress Turbulence model (RST) are applied by using the commercial CFD code STAR-CCM +. For LES and DNS, the open-source CFD code, Nek5000, is used given its outstanding scalability on High Performance Computer (HPC) and high-order technique. The results from RANS simulations are compared with that from LES and DNS for benchmarking. Both macroscale parameters and turbulence statistics, such as velocity magnitude, lateral velocity and turbulence kinetic energy, are presented and analyzed. The results from RANS simulation achieve good agreement with LES and DNS on velocity and turbulence kinetic energy prediction. The RST turbulence model predicts the most similar flow pattern of lateral velocity as compared to LES and DNS. The Lambda-2 (λ2) criterion with a reasonable threshold is used to demonstrate the instantaneous vortices distribution in the involute channel from both LES and DNS calculation. The DNS simulation captures more detailed turbulence especially near the corner, which explains the discrepancy between LES and DNS results near the corner. The normalized RMS error are defined and calculated to assess the performance of those turbulence models. The RST model captures the anisotropic feature of turbulence, which enable it to outperform other turbulence models for predicting the flow behavior in an involute channel. Although some discrepancies are found between LES and DNS results in the corner, the overall deviations between LES and DNS are found to be small. In conclusion, given that the computational cost of DNS calculation is an order of magnitude higher, using LES data for benchmarking RANS model is a cost-effective approach.

DNS↗
DOE OSTI2012

Investigation of the Geometric Accuracy of Proton Beam Irradiation in the Liver

Purpose: To investigate the geometric accuracy of proton beam irradiation to the liver by measuring the change in Hounsfield units (HUs) after irradiation. Methods and Materials: We examined 21 patients with liver tumors who were treated with respiratory-gated proton beam therapy (PBT). The radiation dose was 66 GyE in 12 patients and 72.6 GyE in 9 patients. Image registration and reslicing of the computed tomography (CT) results obtained within 1 month before and 3 months after PBT was performed, referring to the planning CT image. The resliced CT images obtained after PBT were subtracted from the images obtained before PBT. We investigated whether the area of the large HU change was consistent with the high-dose distribution area using the location of the largest change in HU around the tumor (peak) on the subtracted CT image and the 90% dose distribution area of the planning CT image. Results: The number of patients (n = 20) whose left-right peaks were within the 90% dose distribution area was significantly larger than the number of patients whose anterior-posterior peaks and superior-inferior peaks were within the 90% dose distribution area (n = 14, n = 13, p = 0.034, and p = 0.02, respectively). Twelve patients exhibited a peak within the 90% dose distribution area in all directions. Nine of the 11 patients with smaller 90% confidence intervals of the percent normalization of the beam cycle (BC; 90% BC) showed a peak within the 90% dose distribution area in six directions, and this percentage was higher than that among the patients with larger 90% BC (3/10, p = 0.03). Conclusion: The geometric accuracy of proton beam irradiation to the liver was higher in the left-right direction than in the other directions. Patients with an irregular respiratory rhythm have a greater risk of a reduced geometric accuracy of PBT in the liver.

62 RADIOLOGY AND NUCLEAR MEDICINE↗
DOE OSTI2025

Multidimensional Modeling of Mixture Formation in a Hydrogen-Fueled Heavy-Duty Optical Engine With Direct Injection

Hydrogen (H 2 ), as a carbon-free fuel, is considered as one of the most promising solutions to reduce the carbon footprint of hard-to-decarbonize energy and transportation sectors. As such, hydrogen-fueled internal combustion engines (H 2 ICEs) have recently been receiving increasing attention, particularly in applications such as on-road/off-road heavy-duty transport and combined heat and power. The direct injection (DI) of gaseous hydrogen into the combustion chamber offers great potential for achieving high power density and high engine efficiency, while mitigating the risk of backfire and reducing pre-ignition. However, the numerical simulation of H 2 DI system remains a formidable challenge associated with the high computational cost of reproducing compressible supersonic flow and shocks in narrow injector passages and in near-nozzle regions. In general, there is a lack of well-established and validated practices for the modeling of high-pressure H 2 DI in large-bore engines. Here, to this end, this study focuses on computational fluid dynamics (CFD) modeling of the mixture formation process in a heavy-duty optical engine employing a medium-pressure H 2 DI system. Both large eddy simulations (LES) and Reynolds Averaged Navier–Stokes (RANS) simulations are performed and evaluated against optical data. Gaseous hydrogen is injected into the combustion chamber via a centrally located outward opening hollow-cone injector at a pressure of 40 bar. Simulations are carried out for two injection timings, namely, −120 and −60 °CA. The numerical predictions for H 2 distribution in different horizontal and vertical planes during the compression stroke are systematically compared against optical data obtained through planar laser-induced fluorescence (PLIF) measurements. Overall, the LES approach using the Dynamic Structure model is found to have good predictive capabilities for the early jet penetration in terms of length and shape, as well as the later H 2 distributions. However, the unsteady RANS approach with the renormalization group $k - ϵ$ model, which is widely used by industry to model heavy-duty ICEs, significantly underpredicts the H 2 mixing, even at similar mesh resolution to that used in LES. These results indicate that there is a need for the improvement of mixing submodels within the RANS approach when applied to H 2 DI simulations.

LES↗
DOE OSTI2021

Non-Diffusive Volume Advection with A High Order Interface Reconstruction Method

We show that non-diffusive volume advection in two-dimensions is achieved with several benchmark problems using a newly developed high-order volume of fluids (VOF) interface reconstruction method. (1) A new VOF interface reconstruction method using circular/corner facets (linear facets are a degenerate case of arcs). We create a circular interface facet in each mixed zone by matching neighbor volume with a hybrid Newton’s-bisection method and the local solution is final. In the general case, the new VOF interface reconstruction has 3rd order accuracy and can be easily made seamless. The new method addresses intrinsic issues with Young’s method such as gaps between interface facets in the case of a curved interface, and inability to define curvature nor identify corners. (2) A non-diffusive volume advection scheme. In an ALE advection step, a well-defined interface can be carried over through a Lagrange step and used to compute volume distribution into a relaxed mesh. Then, an interface reconstruction step is performed to redefine the interface in the relaxed mesh. We must point out that the interface carried over is also a solution of interface reconstruction because all the volume fractions in the relaxed mesh are naturally matched. We provide an interface tracking method compatible with our reconstruction scheme, where it is granted to use the prior info as an initial guess to capture sub-mesh resolution features. As a result, we are able to treat multiple facets inside a single mixed cell and obtain highly accurate, non-diffusive solution for advection problems with rather coarse meshes. We show our solutions for two-dimensional incompressible flows with two materials with a) the X + O diagonal translation; b) the Zalesak rotational test; and c) the single vortex spiral test.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Rapid Atomic Ordering Transformation toward Intermetallic Nanoparticles

Chemically ordered intermetallic nanoparticles are promising candidates for energy-related applications such as electrocatalysis. However, the synthesis of intermetallics generally requires long annealing (several hours) to achieve the ordered structure, which causes nanoparticles agglomeration and diminished performance, particularly for catalysis. Herein, we demonstrate a new rapid Joule heating approach that can synthesize highly ordered and well-dispersed intermetallic nanoparticles. As a proof-of-concept, we synthesized fully ordered Pd 3 Pb intermetallic nanoparticles that feature small size distribution (~6 nm). Computational analysis of the L1 2 Pd 3 Pb material suggests that this rapid atomic ordering transformation can be attributed to a vacancy-mediated diffusion mechanism. Moreover, the nanoparticles demonstrate excellent electrocatalytic activity and exceptional stability for the oxygen reduction reaction (ORR), retaining >95% of the current density over 10 h of chronoamperometry test with negligible structural and compositional changes. Furthermore, this study demonstrates a new strategy of providing a new direction for intermetallic synthesis and catalyst discovery.

25 ENERGY STORAGE↗
DOE OSTI2018

User-level File Systems Specialized for HPC Workloads Year-End Report

High-performance computing (HPC) clusters attract the attention of Deep Learning (DL) training users due to the clusters' powerful computation capabilities. While there have been many existing efforts to enable deep neural works to leverage the powerful CPU and GPU processors from leadership high-performance computing (HPC) systems, large-scale deep learning with larger datasets requires efficient I/O support from the underlying file and storage systems. As some current and upcoming HPC clusters have large on-node memory or are equipped with NVMe SSD on compute nodes, more distributed DL trainings are considered to leverage those storages to store dataset for efficient dataset access. Our project goal is to design a specialized DL-oriented file system to improve datasets loading performance of any DL training applications. Over the past year, the team at Florida State University has performed research activities in three different aspects, including 1) completing a specialized memory-based I/O framework (DeepIO) for improving dataset loading performance on DL applications; 2) proposing a more generalized file system (DLFS) for DL applications on node-local SSDs; 3) completing the BeeGFS performance evaluation project for Deep Neural Networks.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2019

Sparse Recovery for Scientific Data

We have built build and analyzed fundamental mathematical tools that enable sparse recovery of scientific data for high performance computing, experimental facilities and sensor measurement. The proposed work will impact many areas of research across the DOE. We have designed, developed and analyzed fast scalable algorithms for the sparse recovery of scientific data sets that may be extremely large, generated very quickly and distributed across multiple storage devices. The algorithms we developed will be used to process the copious amounts of scientific data generated by the DOE supercomputer simulations and by DOE experimental and observational facilities. They leverage the fact that while the amount of data produced can be extremely large, the actual quantities of interest (QoI's) are often sparse, either in space time, or in some feature space. Mathematically, notions of compressive sensing, including l1 regularized and total variation regularized optimization have played an important part in this effort.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Predicting Mechanical Properties from Microstructure Images in Fiber-Reinforced Polymers Using Convolutional Neural Networks

Evaluating the mechanical response of fiber-reinforced composites can be extremely time-consuming and expensive. Machine learning (ML) techniques offer a means for faster predictions via models trained on existing input–output pairs and have exhibited success in composite research. This paper explores a fully convolutional neural network modified from StressNet, which was originally used for linear elastic materials, and extended here for a non-linear finite element (FE) simulation to predict the stress field in 2D slices of segmented tomography images of a fiber-reinforced polymer specimen. The network was trained and evaluated on data generated from the FE simulations of the exact microstructure. The testing results show that the trained network accurately captures the characteristics of the stress distribution, especially on fibers, solely from the segmented microstructure images. The trained model can make predictions within seconds in a single forward pass on an ordinary laptop, given the input microstructure, compared to 92.5 h to run the full FE simulation on a high-performance computing cluster. These results show promise in using ML techniques to conduct fast structural analysis for fiber-reinforced composites and suggest a corollary that the trained model can be used to identify the location of potential damage sites in fiber-reinforced polymers.

Sun, Yixuan (ORCID:0000000311093380)↗
DOE OSTI2023

Final Report on Development of Microwave Resonant Cavity Transducer for Fluid Flow Sensing: Development of Microwave Cavity Flow Meter for Advanced Reactor High Temperature Fluids

We are investigating a microwave resonant cavity transducer for flow sensing in the vessel of a high temperature fluid advanced reactor (AR), such as a molten salt cooled reactor (MSCR) or a sodium fast reactor (SFR). This transducer is a hollow metallic cylindrical cavity, with one of the flat walls of the cylinder flexible enough to undergo microscopic deflection due to dynamic fluid pressure. Membrane deflection leads to a shift in the microwave resonant frequency, which can be detected with a spectrum analyzer. We have developed a continuum electromechanics model of the microwave resonant cavity transducer performance. The deflection of the membrane is calculated using equations from the theory of plates and shells for deflection of radially constrained thin circular disk under uniform load. The microwave frequency shift is calculated using the equations from the microwave cavity perturbation theory. The transducer model was validated with proof-of-concept experimental data of water flow sensing with a Brass 360 right circular cylinder with 22.2mm diameter and 203µm thick wall. The cavity was excited through a subwavelength hole in TE 011 mode with resonant frequency f ≈ 17.8GHz. Subsequently, we have performed a preliminary proof-of-principle test of flow sensing in high temperature liquid sodium in environment. For this test, we have developed a cylindrical resonator with the same dimensions as for the Brass 360 cavity. The resonator for high temperature liquid sodium test was machined from stainless steel 316 and electroplated with silver on the interior surfaces. We have also developed and insertion probe consisting of a 50cm WR-42 brass waveguide enclosed in a protective SS316 tube. The cavity was excited through a subwavelength hole on the side of the wall of the cylinder in the TE 011 mode with resonant frequency f ≈ 17.8GHz. The liquid sodium setup consists of a cylindrical vessel with a center feed line, where a transducer inserted through the top cap of the vessel measures velocity of the impinging liquid jet. The sodium flow sensing study was performed in a liquid sodium vessel at 340°C temperature and ambient pressure. The flow rate was changed by varying sodium pump power, and a frequency shift of several hundred kHz was observed. The transducer was removed after 70 days of testing. No visual damage was observed, and the electromagnetic response remained the same. As a calibration experiment, we have assembled a water vessel with a center feed, and with dimensions similar to those of the liquid sodium setup. Water flow sensing was performed before and after liquid sodium test. The responses in both cases are similar and agree with COMSOL computer simulations. Because long-term (multi-year) experimental tests of transducer resilience to harsh environment are not practical, we have developed a probabilistic model of creep to estimate transducer resilience to the harsh environment. The probabilistic model considers diffusion creep under the condition of high temperature and low stress, where the stress and temperature are allowed to be random variables with Gaussian distributions. Using the probabilistic model, we estimate inelastic membrane deflections due to creep for several temperature ranges. We conclude that for temperatures less than 650°C, creep has negligible long-term effect on the transducer performance.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2017

NPTFit: A Code Package for Non-Poissonian Template Fitting

We present NPTFit, an open-source code package, written in Python and Cython, for performing non-Poissonian template fits (NPTFs). The NPTF is a recently developed statistical procedure for characterizing the contribution of unresolved point sources (PSs) to astrophysical data sets. The NPTF was first applied to Fermi gamma-ray data to provide evidence that the excess of ∼GeV gamma-rays observed in the inner regions of the Milky Way likely arises from a population of sub-threshold point sources, and the NPTF has since found additional applications studying sub-threshold extragalactic sources at high Galactic latitudes. The NPTF generalizes traditional astrophysical template fits to allow for the ability to search for populations of unresolved PSs that may follow a given spatial distribution. NPTFit builds upon the framework of the fluctuation analyses developed in X-ray astronomy, thus it likely has applications beyond those demonstrated with gamma-ray data. The NPTFit package utilizes novel computational methods to perform the NPTF efficiently. The code is available at http://github.com/bsafdi/NPTFit and up-to-date and extensive documentation may be found at http://nptfit.readthedocs.io.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2022

Accelerated Coupled Monte Carlo-Thermal Hydraulic Calculations using a Hybrid GTF-Diffusion-based Prediction Block: First Results

Accurate predictions of spatial power and temperature distributions require the coupling of a neutron transport solver with a thermal-hydraulic (TH) feedback. Nowadays, Monte Carlo (MC) codes are widely coupled to TH solvers, typically via a Picard iteration (PI) method, due to the higher fidelity that such frameworks can produce. To speed up a PI, a prediction step can produce an improved initial guess for a source distribution and feed it to the MC code. Recent work [1, 2] investigated a prediction step that uses generalized transfer functions (GTFs) to predict the macroscopic cross sections’ variations following a perturbation in TH properties, such as coolant density. The previous method also relied on first order perturbation (FOP) theory to predict perturbed power profiles, rather than using an expensive MC iterate. The implemented FOP method relied on generating a fission matrix from which the forward and adjoint eigenmodes were extracted and later used to for power calculations. The generation of the fission matrix can introduce a significant computational overhead, therefore undermining the performance of the proposed hybrid technique when applied to high-dimensional problems, e.g., full core calculations. This work attempts to improve the GTF-FOP prediction step by replacing the FOP solver with a nodal diffusion solver, thus eliminating the need to calculate a fission matrix. The GTF-diffusion step was tested for various moderator density perturbations. In each case, the predicted power distribution showed good agreement with the reference case. The latter is attributed to the generally good prediction of most spatially distributed macroscopic cross sections, except the transport cross section, which will become the focus of future work.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2021

Machine Learning-enabled Scalable Performance Prediction of Scientific Codes

Hardware architectures become increasingly complex as the compute capabilities grow to exascale. Here, we present the Analytical Memory Model with Pipelines (AMMP) of the Performance Prediction Toolkit (PPT). PPT-AMMP takes high-level source code and hardware architecture parameters as input and predicts runtime of that code on the target hardware platform, which is defined in the input parameters. PPT-AMMP transforms the code to an (architecture-independent) intermediate representation, then (i) analyzes the basic block structure of the code, (ii) processes architecture-independent virtual memory access patterns that it uses to build memory reuse distance distribution models for each basic block, and (iii) runs detailed basic-block level simulations to determine hardware pipeline usage. PPT-AMMP uses machine learning and regression techniques to build the prediction models based on small instances of the input code, then integrates into a higher-order discrete-event simulation model of PPT running on Simian PDES engine. We validate PPT-AMMP on four standard computational physics benchmarks and present a use case of hardware parameter sensitivity analysis to identify bottleneck hardware resources on different code inputs. We further extend PPT-AMMP to predict the performance of a scientific application code, namely, the radiation transport mini-app SNAP. To this end, we analyze multi-variate regression models that accurately predict the reuse profiles and the basic block counts. We validate predicted SNAP runtimes against actual measured times.

97 MATHEMATICS AND COMPUTING↗