GPGPU Research Articles

HiPACE++: A portable, 3D quasi-static particle-in-cell code

Modeling plasma accelerators is a computationally challenging task and the quasi-static particle-in-cell algorithm is a method of choice in a wide range of situations. In this work, we present the first performance-portable, quasi-static, three-dimensional particle-in-cell code HiPACE++. By decomposing all the computation of a 3D domain in successive 2D transverse operations and choosing appropriate memory management, HiPACE++ demonstrates orders-of-magnitude speedups on modern scientific GPUs over CPU-only implementations. The 2D transverse operations are performed on a single GPU, avoiding time-consuming communications. The longitudinal parallelization is done through temporal domain decomposition, enabling near-optimal strong scaling from 1 to 512 GPUs. HiPACE++ is a modular, open-source code enabling efficient modeling of plasma accelerators from laptops to state-of-the-art supercomputers. Program summaryProgram Title: HiPACE++CPC Library link to program files:https://doi.org/10.17632/zh3rc7hvrm.1Developer's repository link:HiPACE++ GitHub repositoryLicensing provisions: BSD 3-clauseProgramming language: C++Nature of problem: Modeling plasma accelerators is a computationally challenging task requiring nanometer-scale resolutions over meter-scale propagation distances. The quasi-static particle-in-cell method enables high-fidelity simulations of this strongly non-linear process, but these simulations can be very expensive.Solution method: The quasi-static particle-in-cell algorithm is modified to enable efficient utilization of accelerated hardware, in particular with GPU computing, reducing the cost of simulations by orders of magnitude. A novel longitudinal parallelization enables excellent strong scaling of this method up to hundreds of GPUs.Reference:10.5281/zenodo.5358483

The Application of AI Technology in GPU Scheduling Algorithm Optimization

With the rapid development of integrated circuit technology, GPU computing capabilities continue to improve. Due to the continuous improvement and improvement of GPU programming capabilities, functions, and performance, GPUs have been widely used in the field of high-tech general-purpose computers. This article is aimed at studying the optimization of GPU scheduling algorithm based on AI technology. Through a combination of theoretical analysis and simulation experiments, the concepts of artificial intelligence technology and GPU scheduling are explained, and the impact of GPU architecture and GPGPU load on the energy efficiency of GPGPU is explained. On the basis of comprehensive analysis of GPU cluster characteristics, a new GA-TP scheduling algorithm based on genetic algorithm was designed, and based on the energy efficiency of the cluster, a simulation verification platform was built for the accuracy of simulation. Experimental results show that the acceleration rate of the GA-TP algorithm is significantly lower than that of the HEFT algorithm, the average acceleration rate is reduced by nearly 25%, and the scheduling efficiency of the GA-TP algorithm is higher.

Fast Study Quadric Interpolation in the Conformal Geometric Algebra Framework

Interpolating trajectories of points and geometric entities is an important problem for kinematics. To describe these trajectories, several algorithms have been proposed using matrices, quaternions, dual-quaternions, and the Study quadric; the last one allows the embedding of motors as 8D vectors into projective space P7, where the interpolation of rotations and translations becomes a linear problem. Furthermore, conformal geometric algebra (CGA) is an effective and intuitive framework for representing and manipulating geometric entities in Euclidean spaces, and it allows the use of quaternions and dual-quaternions formulated as Motors. In this paper, a new methodology for accelerating the Study quadric Interpolation based on Conformal Geometric Algebra is presented. This methodology uses General Purpose Graphics Processing Units (GPUs) and it is applied for medical robotics, but it can also be extended to other areas such as aeronautics, robotics, and graphics processing.

Extremely large deformation of tunnel induced by rock mass fracture using GPGPU parallel FDEM

AbstractIt is essential to better understand the large deformation behavior of deep tunnels in weak rock mass to improve the long‐term stability of tunnel surrounding rock mass and design effective supports for squeezing tunnels. This paper introduces two representative large deformation engineering examples to study squeezing behavior. First, new algorithms of contact search, determination of actual contact, and calculation law of contact pairs are proposed for general‐purpose computing on graphics processing units (GPGPUs), by improving the original serial algorithms and developing a compute unified device architecture (CUDA) parallel program, to satisfy the large‐scale rock fracture process simulation. Then, tunnels excavation and rock reinforcement effect including the combination of U‐shaped steel and steel fabric, spray concrete (SMC), and grouting were modeled. The verification, done by comparing the field observation data and the FDEM simulation results of progressive fracture and the tunnel wall convergence displacement, shows that stress remarkable rebalancing after excavation because of high in situ stress, low rock mass strength, and rock mass long‐term strength reduction properties, thus, the rock mass fracture progressively and the rock blocks move continuously, leading to the rock mass volume increase significantly. The result demonstrates one of the root causes of squeezing. To effectively reduce the tunnel wall displacement and maintain tunnel stability, support with higher stiffness or multifactorial combined support methods and grouting with a better bonding effect should be used together. The findings provide insights into the progress of tunnel squeezing and the basis for support design in the squeezing tunnel.

Multi-task hourglass network for online automatic diagnosis of developmental dysplasia of the hip.

Developmental dysplasia of the hip (DDH) is one of the most common diseases in children. Due to the experience-requiring medical image analysis work, online automatic diagnosis of DDH has intrigued the researchers. Traditional implementation of online diagnosis faces challenges with reliability and interpretability. In this paper, we establish an online diagnosis tool based on a multi-task hourglass network, which can accurately extract landmarks to detect the extent of hip dislocation and predict the age of the femoral head. Our method utilizes a multi-task hourglass network, which trains an encoder-decoder network to regress the landmarks and predict the developmental age for online DDH diagnosis. With the support of precise image analysis and fast GPU computing, our method can help overcome the shortage of medical resources and enable telehealth for DDH diagnosis. Applying this approach to a dataset of DDH X-ray images, we demonstrate 4.64 mean pixel error of landmark detection compared to the results of human experts. Moreover, we can improve the accuracy of the age prediction of femoral heads to 89%. Our online automatic diagnosis system has provided service to 112 patients, and the results demonstrate the effectiveness of our method.

An efficient and robust GPGPU-parallelized contact algorithm for the combined finite-discrete element method

In this study, an efficient and robust GPGPU (general purpose graphic processing unit)-parallelized algorithm is proposed for contact force calculation for the three-dimensional combined finite-discrete element method (3D FDEM). The contact force model is energy-conserving and avoids the issues that the mesh-dependence of contact force in the original contact algorithm proposed by Munjiza. The contact force calculation is based on the determination of the geometrical features of the overlapped region between two particles, which can be easily obtained by performing the face–particle intersection calculations consecutively. The contact damping and contact friction are also implemented. Based on Compute Unified Device Architecture (CUDA), the proposed contact algorithm can be parallelized with much less thread unbalance and register usage than the existing energy-conserving contact algorithms due to the simplified computational process. Several numerical tests are performed to validate the efficiency and effectiveness of the proposed contact algorithm. The contact friction model is validated against theoretical solutions with a block sliding test, and the results show that the proposed contact algorithm outperforms the original contact algorithm in both normal and friction force evaluation. For both quasi-static and dynamic scenarios, the simulated results are in good agreement with those generated by the original contact algorithm, as well as the experimental measurements. The computational efficiency tests show that the computation time is linearly proportional to the number of potential contact pairs, and the speedup ratio of the parallelized version of proposed contact algorithm relative to the serial version of original contact algorithm can reach up to 565.1 (Nvidia Quadro GP100), which indicates the proposed contact algorithm is capable to be employed in the simulation of large-scale problems. Besides, the proposed contact algorithm can be applied to the cases where the polyhedrons with more than four vertexes are employed.

PKNN-MIFS: A Parallel KNN Classifier over an Optimal Subset of Features

Feature selection help select an optimal subset of features from a large feature space to achieve better classification performance. The performance of KNN classifier can be improved significantly using an appropriate subset of features from a large feature space. Recent development in General Purpose Graphics Processing Units (GPGPU) has provided us a low cost yet high performance computing support for wide range of applications. This paper presents a parallel KNN classifier powered by a mutual information based feature selection called PKNN-MIFS for effective classification of real life data. It selects an optimal subset of features from the original feature set by exploiting the mutual information concept for the estimation of feature-class and feature-feature relevance. It selects a non-redundant feature by giving higher priority on feature-class relevance. The performance of the proposed PKNN-MIFS has been evaluated over several datasets and has been found to be superior to its closed counterpart.

Survey: Tensorflow in Machine Learning

TensorFlow(TF) is a large-scale ML system that can be used in many different situations. Data-flow graphs are used by TF to define operations, shared state, and computations that change that state. It distributes the nodes of a Data-flow graph across multiple machines in a cluster, as well as diverse processing units within a single system, such as multicore CPUs, general-purpose GPUs and Tensor Processing Units, which are custom-designed ASICs (TPUs). Developers may use TF to try out new training techniques and optimizations. It covers a broad applications range, which focus on inference and DNN training. It is utilised in production by several Google services, and it has since been released as an open-source project and is broadly used in ML research. We review TensorFlow and place it in the context of modern deep learning techniques and technologies in this paper. We go over the fundamental computational concepts and how they work. The distributed execution concept, as well as its programming interface as accompanying visualisation toolkit. The remainder of this paper will provide a comprehensive overview of TensorFlow and place it in the context of machine learning today

BINER: A low-cost biomedical named entity recognition

A primary focus of the healthcare industry is to improve patient experience and quality of service. Practitioners and health workers are generating large volumes of text that are captured in Electronic Medical Records, clinical reports, and publications. Additionally, patients post millions of comments on social media related to healthcare, on diverse topics such as hospital services, disease symptoms, and drugs effects. Unifying various data sources can guide physicians and healthcare workers to avoid unnecessary, irrelevant information and expedite access to helpful information. The main challenge to creating Biomedical Natural Language Understanding is the lack of standard datasets and the extensive computational resources needed to develop different models. This paper proposes a model trained on low-tier GPU computers, producing comparable results to larger models like BioBERT. We propose BINER, a Biomedical Named Entity Recognition architecture using limited data and computational resources.

Parallel Logic Programming: A Sequel

AbstractMulti-core and highly connected architectures have become ubiquitous, and this has brought renewed interest in language-based approaches to the exploitation of parallelism. Since its inception, logic programming has been recognized as a programming paradigm with great potential for automated exploitation of parallelism. The comprehensive survey of the first twenty years of research in parallel logic programming, published in 2001, has served since as a fundamental reference to researchers and developers. The contents are quite valid today, but at the same time the field has continued evolving at a fast pace in the years that have followed. Many of these achievements and ongoing research have been driven by the rapid pace of technological innovation, that has led to advances such as very large clusters, the wide diffusion of multi-core processors, the game-changing role of general-purpose graphic processing units, and the ubiquitous adoption of cloud computing. This has been paralleled by significant advances within logic programming, such as tabling, more powerful static analysis and verification, the rapid growth of Answer Set Programming, and in general, more mature implementations and systems. This survey provides a review of the research in parallel logic programming covering the period since 2001, thus providing a natural continuation of the previous survey. In order to keep the survey self-contained, it restricts its attention to parallelization of the major logic programming languages (Prolog, Datalog, Answer Set Programming) and with an emphasis on automated parallelization and preservation of the sequential observable semantics of such languages. The goal of the survey is to serve not only as a reference for researchers and developers of logic programming systems but also as engaging reading for anyone interested in logic and as a useful source for researchers in parallel systems outside logic programming.

3D Numerical Analysis Method for Simulating Collapse Behavior of RC Structures by Hybrid FEM/DEM

Recent years have seen an increase in demand for the demolition of obsolete and potentially hazardous structures, including reinforced concrete (RC) structures, using blasting techniques. However, because the risk of failure is significantly higher when applying blasting to demolish RC structures than mechanical dismantling, it is critical to achieve the optimal demolition design and conditions using blasting by taking into account the major factors affecting a structure’s demolition. To this end, numerical analysis techniques have frequently been used to simulate the progressive failure resulting in the collapse of structures. In this study, the three-dimensional (3D) combined finite discrete element method (FDEM), which is accelerated by a parallel computation technique incorporating a general-purpose graphics processing unit (GPGPU), was coupled with the one-dimensional (1D) reinforcing bar (rebar) model as a numerical simulation tool for simulating the process of RC structure demolition by blasting. Three-point bending tests on the RC beams were simulated to validate the developed 3D FDEM code, including the calibration of 3D FDEM input parameters to simulate the concrete fracture in the RC beam accurately. The effect of the elements size for the concrete part on the RC beam’s fracture process was also discussed. Then, the developed 3D FDEM code was used to model the blasting demolition of a small-scale RC structure. The numerical simulation results for the progressive collapse of the RC structure were compared to the actual experimental results and found to be highly consistent.

Evaluation of NVIDIA Xavier NX Platform for Real-Time Image Processing for Plasma Diagnostics

Machine protection is a core task of real-time image diagnostics aiming for steady-state operation in nuclear fusion devices. The paper evaluates the applicability of the newest low-power NVIDIA Jetson Xavier NX platform for image plasma diagnostics. This embedded NVIDIA Tegra System-on-a-Chip (SoC) integrates a Graphics Processing Unit (GPU) and Central Processing Unit (CPU) on a single chip. The hardware differences and features compared to the previous NVIDIA Jetson TX2 are signified. Implemented algorithms detect thermal events in real-time, utilising the high parallelism provided by the embedded General-Purpose computing on Graphics Processing Units (GPGPU). The performance and accuracy are evaluated on the experimental data from the Wendelstein 7-X (W7-X) stellarator. Strike-line and reflection events are primarily investigated, yet benchmarks for overload hotspots, surface layers and visualisation algorithms are also included. Their detection might allow for automating real-time risk evaluation incorporated in the divertor protection system in W7-X. For the first time, the paper demonstrates the feasibility of complex real-time image processing in nuclear fusion applications on low-power embedded devices. Moreover, GPU-accelerated reference processing pipelines yielding higher accuracy compared to the literature results are proposed, and remarkable performance improvement resulting from the upgrade to the Xavier NX platform is attained.

GPU-Accelerated Online Short-Circuit Interrupting Capacity Scan Based on Unified Modeling

With the increasing of transmission density and short-circuit current level, interruption capacity scanning has become an important approach for the online assessment of the circuit breaker. The short-circuit calculation scale and the expected failures significantly increase under the demand of unified analysis of the whole grid, which cause the considerable challenges for online assessment. In this study, we propose a GPU-accelerated short-circuit interruption capacity scanning algorithm for two different unified models, that is, GPU-based 2D and 3D expansion models. For both models, the redundant configuration of the fault network matrices is designed to reduce the cost of GPU computation process. Then, the batch node voltage analysis and the batch branch current calculation are derived using the considered model. In addition, the differences between two models are discussed in detail. Simulations show that the proposed algorithm with GPU-based 3D-expand model has a 3.45× speedup compared with the traditional algorithm on a 9241-bus system with 8-core CPU. Therefore, the proposed algorithms have the potential to handle the full short-circuit current scan under the unified analysis scenario.

A ML-based resource utilization OpenCL GPU-kernel fusion model

Massive data parallelism can be achieved by using general-purpose graphics processing units (GPGPU) with the help of the OpenCL framework. When smaller data with higher GPU memory is executed, it results in a low resource utilization ratio and energy inefficiencies. Up until now, there is no existing model to share GPU for further execution. In addition, if the kernel pair requires the same computation resource, then kernel merging also results in a significant increase in execution time. Therefore, optimal device selection, as well as kernel merging, can significantly speed up the execution performance for a batch of jobs. This paper proposes a kernel merging method that leads to high GPU occupancy. As a result, it reduces execution time and increases GPU utilization. Additionally, a machine learning (ML)-based GPU sharing mechanism is presented to select pairs of kernels in OpenCL frameworks. The model first selects suitable architecture for the jobs and then merges GPU kernels for better resource utilization. From all the GPU candidates, the optimal pair of the kernel concerning data size is selected. The experimental results show that the developed model can achieve 0.91 F1-measure for device selection and 0.98 for the scheduling scheme of kernel merging.

Accelerating CFD simulation with high order finite difference method on curvilinear coordinates for modern GPU clusters

A high fidelity flow simulation for complex geometries for high Reynolds number (Re) flow is still very challenging, requiring a more powerful HPC system. However, the development of HPC with traditional CPU architecture suffers bottlenecks due to its high power consumption and technical difficulties. Heterogeneous architecture computation is raised to be a promising solution to the challenges of HPC development. GPU accelerating technology has been utilized in low order scheme CFD solvers on the structured grid and high order scheme solvers on unstructured meshes. The high-order finite difference methods on structured grids possess many advantages, e.g., high efficiency, robustness, and low storage. However, the strong dependence among points for a high-order finite difference scheme still limits its application on the GPU platform. In the present work, we propose a set of hardware-aware technology to optimize data transfer efficiency between CPU and GPU, as well as communication efficiency among GPUs. An in-house multi-block structured CFD solver with high order finite difference methods on curvilinear coordinates is ported onto the GPU platform and obtains satisfying performance with a speedup maximum of around 2000x over a single CPU core. This work provides an efficient solution to apply GPU computing in CFD simulation with specific high order finite difference methods on current GPU heterogeneous computers. The test shows that significant accelerating effects can be achieved for different GPUs.

General-purpose GPU hashing data structures and their application in accelerated genomics

A broad variety of applications relies on associative data structures that exclusively support insert, retrieve, and delete operations. Hash maps represent such a class of effective dictionary implementations. Properties such as amortized constant time complexity for these table operations as well as a compact memory layout make them versatile data structures with manifold applications in data analytics and artificial intelligence. The rapidly growing amount of data emerging in many scientific fields can often only be tackled with modern massively parallel accelerators such as GPUs. Numerous GPU hash table implementations have been proposed over the recent years. However, most of these implementations lack flexibility in order to be used in existing analytics pipelines or suffer from significant performance degradation for certain application scenarios. As a more recent approach, the WarpCore framework aims to alleviate these aforementioned restrictions by placing a focus on both versatility and performance. In this work we reflect the key concepts of the WarpCore library and provide an extensive performance evaluation against the state-of-the-art. We further explore how WarpCore can be used for accelerating two bioinformatics applications (metagenomic classification and k-mer counting) with significant speedups.

Tensorox: Accelerating GPU Applications via Neural Approximation on Unused Tensor Cores

Driven by the demands of deep learning, many hardware accelerators, including GPUs, have begun to include specialized tensor processing units to accelerate matrix operations. However, general-purpose GPU applications that have little or no large dense matrix operations cannot benefit from these tensor units. This article proposes Tensorox, a framework that exploits the half-precision tensor cores available on recent GPUs for approximable, non deep learning applications. In essence, a shallow neural network is trained based on the input-output mapping of the function to be approximated. The key innovation in our implementation is the use of the small and dimension-restricted tensor operations in Nvidia GPUs to run multiple instances of the approximation neural network in parallel. With the proper scaling and training methods, our approximation yielded an overall accuracy that is higher than naïvely running the original programs with half-precision. Furthermore, Tensorox allows for the runtime adjustment of the degree of approximation. For the 10 benchmarks we tested, we achieved speedups from 2× to 112× compared to the original in single precision floating point, while maintaining the error caused by the approximation to below 10 percent in most applications.

Parallel CPU–GPU computing technique for discrete element method

AbstractThe efficiency of the simulations with the discrete element method (DEM) is significantly improved using a novel computational strategy. The new method is developed with a focus on platforms equipped with multi‐core central processing units (CPU) and general‐purpose graphics processing units (GPU). The DEM calculations are performed in parallel on the CPU and on the GPU using pre‐calculated Verlet lists with a posteriori analysis of their consistency. The operations related to the search for possible contacts are performed on the CPU, whereas the processing of interactions, and integration of motion, are executed on the GPU. Performance analysis done for various types of tasks has shown that the new method allows to significantly decrease the average computational time and to utilize available computational resources more efficiently compared to the sequential CPU–GPU execution mode. Furthermore, due to more efficient calculations, the overall energy requirement for the proposed strategy does not exceed the demand for conventional sequential CPU–GPU computations.

Efficient Machine Learning execution with Near-Data Processing

A myriad of Machine Learning (ML) algorithms has emerged as a basis for many applications due to the facility in obtaining satisfactory solutions to a wide range of problems. Programs such as K-Nearest Neighbors (KNN), Multi-layer Perceptron (MLP), and Convolutional Neural Network (CNN) are commonly applied on Artificial Intelligence (AI) to process and analyze the ever-increasing amount of data. Nowadays, multi-core general-purpose systems are adopted due to their high processing capacity. However, energy consumption tends to scale together with the number of used cores. In order to achieve performance and energy efficiency, many Near-Data Processing (NDP) architectures have been proposed mainly tackling data movement reduction by placing computing units as close as possible to the data, such as specific accelerators, full-stack General Purpose Processor (GPP) and Graphics Processing Unit (GPU), and vector units. In this work, we present the benefits of exploring Vector-In-Memory Architecture (VIMA), a general-purpose vector-based NDP architecture for varied ML algorithms. Our work directly compares VIMA and x86 multi-core AVX-512 GPP. The presented approach can overcome the x86 single-core in up to 11.3× while improving energy efficiency by up to 8×. Also, our study shows that nearly 16 cores are necessary to match the NDP’s single-thread performance for KNN and MLP algorithms, while it is necessary 32 cores for convolution. Nevertheless, VIMA still overcomes x86 32 cores by 2.1× on average when considering energy results.

Accelerating an iterative eigensolver for nuclear structure configuration interaction calculations on GPUs using OpenACC

To accelerate the solution of large eigenvalue problems arising from many-body calculations in nuclear physics on distributed-memory parallel systems equipped with general-purpose Graphic Processing Units (GPUs), we modified a previously developed hybrid MPI/OpenMP implementation of an eigensolver written in Fortran 90 by using an OpenACC directives based programming model. Such an approach requires making minimal changes to the original code and enables a smooth migration of large-scale nuclear structure simulations from a distributed-memory many-core CPU system to a distributed GPU system. However, in order to make the OpenACC based eigensolver run efficiently on GPUs, we need to take into account the architectural differences between a many-core CPU and a GPU device. Consequently, the optimal way to insert OpenACC directives may be different from the original way of inserting OpenMP directives. We point out these differences in the implementation of sparse matrix–vector multiplications with multiple vectors, which constitutes the main cost of the eigensolver, as well as other differences in the preconditioning step and dense linear algebra operations. We compare the performance of the OpenACC based implementation executed on multiple GPUs with the performance on distributed-memory many-core CPUs, and demonstrate significant speedup achieved on GPUs compared to the on-node performance of a many-core CPU. We also show that the overall performance improvement of the eigensolver on multiple GPUs is more modest due to the communication overhead among different MPI ranks.