GPGPU Research Articles

GPGPU를 활용한 OpenFOAM 기반 해석자 성능 분석

Two benchmark tests were carried out to analyze the performance of OpenFOAM-based CFD solvers using General-Purpose computing on Graphics Processing Units(GPGPU). In the present study, RapidCFD, which is an implementation of OpenFOAM capable of running most of the functions of OpenFOAM on GPUs, was used to apply GPGPU to OpenFOAM. The numerical simulations of 1) 3D lid-driven cavity incompressible flows and 2) steady flows around a motorbike were conducted on two kinds of CPU, single-GPU, and multi-GPU systems, and the computational times were analyzed. For the test of cavity flows, as the number of cells increased, the performance and the scalability of GPGPU were improved. When the number of cells was 2503, a system with 8-GPUs showed the highest performance with 42 times of speedup over a CPU system. For the test of flows around a motorbike, a system with 8-GPUs showed the highest performance with 20 times of speedup over a CPU system. For both single precision and double precision calculations, the performance improvements using GPGPU were efficient. The results demonstrate that GPGPU would be more efficient than computing on CPUs when computing large-scale flows and practical problems that require massive parallelism.

Out-of-core GPU 2D-shift-FFT algorithm for ultra-high-resolution hologram generation.

We propose a novel out-of-core GPU algorithm for 2D-Shift-FFT (i.e., 2D-FFT with FFT-shift) to generate ultra-high-resolution holograms. Generating an ultra-high-resolution hologram requires a large complex matrix (e.g., 100K2) with a size that typically exceeds GPU memory. To handle such a large-scale hologram plane with limited GPU memory, we employ a 1D-FFT based 2D-FFT computation method. We transpose the column data to have a continuous memory layout to improve the column-wise 1D-FFT stage performance in both the data communication and GPU computation. We also combine the FFT-shift and transposition steps to reduce and hide the workload. To maximize the GPU utilization efficiency, we exploit the concurrent execution ability of recent heterogeneous computing systems. We also further optimize our method's performance with our cache-friendly chunk generation algorithm and pinned-memory buffer approach. We tested our method on three computing systems having different GPUs and various sizes of complex matrices. Compared to the conventional implementation based on the state-of-the-art GPU FFT library (i.e., cuFFT), our method achieved up to 3.24 and 3.06 times higher performance for a large-scale complex matrix in single- and double-precision cases, respectively. To assess the benefits offered by the proposed approach in an actual application, we applied our method to the layer-based CGH process. As a result, it reduced the time required to generate an ultra-high-resolution hologram (e.g., 100K2) up to 28% compared to the use of the conventional algorithm. These results demonstrate the efficiency and usefulness of our method.

A unified schedule policy of distributed machine learning framework for CPU-GPU cluster

With the widespread using of GPU hardware facilities, more and more distributed machine learning applications have begun to use CPU-GPU hybrid cluster resources to improve the efficiency of algorithms. However, the existing distributed machine learning scheduling framework either only considers task scheduling on CPU resources or only considers task scheduling on GPU resources. Even considering the difference between CPU and GPU resources, it is difficult to improve the resource usage of the entire system. In other words, the key challenge in using CPU-GPU clusters for distributed machine learning jobs is how to efficiently schedule tasks in the job. In the full paper, we propose a CPU-GPU hybrid cluster schedule framework in detail. First, according to the different characteristics of the computing power of the CPU and the computing power of the GPU, the data is divided into data fragments of different sizes to adapt to CPU and GPU computing resources. Second, the paper introduces the task scheduling method under the CPU-GPU hybrid. Finally, the proposed method is verified at the end of the paper. After our verification for K-Means, using the CPU-GPU hybrid computing framework can increase the performance of K-Means by about 1.5 times. As the number of GPUs increases, the performance of K-Means can be significantly improved.

A Taxonomy of Modern GPGPU Programming Methods: On the Benefits of a Unified Specification

Several Application Programming Interfaces (APIs) and frameworks have been proposed to simplify the development of General-Purpose GPU (GPGPU) applications. GPGPU application development typically involves specific customization for the target operating systems and hardware devices. The effort to port applications from one API to the other (or to develop multi-target applications) is complicated by the availability of a plethora of specifications, which in essence offers very similar underlying functionality. In this work we provide an in-depth study of six state-of-the-art GPGPU APIs. From these we derive a taxonomy of the common semantics and propose a unified specification. We describe a methodology to translate this unified specification into different target APIs. This simplifies cross-platform application development and provides a clean framework for benchmarking. Our proposed unified specification is called GUST (GPGPU Unified Specification and Translation) and it captures common functionality found in compute-only APIs (e.g., CUDA and OpenCL), in the compute pipeline of traditional graphic-oriented APIs (e.g., OpenGL and Direct3D11) and in last-generation bare-metal APIs (e.g., Vulkan and Direct3D12). The proposed translation methodology solves differences between specific APIs in a transparent manner, without hiding available tuning knobs for compute kernel optimizations and fostering best programming practices in a simple manner.

A GPU-Accelerated Mixed-Precision WENO Method for Extremal Black Hole and Gravitational Wave Physics Computations

We develop and use a novel mixed-precision weighted essentially non-oscillatory (WENO) method for solving the Teukolsky equation, which arises when modeling perturbations of Kerr black holes. We show that WENO methods outperform higher-order finite-difference methods, standard in the discretization of the Teukolsky equation, due to the need to add dissipation for stability purposes in the latter. In particular, as the WENO scheme uses no additional dissipation, it is well suited for scenarios requiring long-time evolution such as the study of price tails and gravitational wave emission from extreme mass ratio binaries. In the mixed-precision approach, the expensive computation of the WENO weights is performed in reduced floating-point precision that results in a significant speedup factor of \(\approx 3.3\). In addition, we use state-of-the-art Nvidia general-purpose graphics processing units and cluster parallelism to further accelerate the WENO computations. Our optimized WENO solver can be used to quickly generate accurate results of significance in the field of black hole and gravitational wave physics. We apply our solver to study the behavior of the Aretakis charge—a conserved quantity, that if detected by a gravitational wave observatory like LIGO/Virgo would prove the existence of extremal black holes.

Simulation of 3D centimeter-scale continuum tumor growth at sub-millimeter resolution via distributed computing

Simulation of cm-scale tumor growth has generally been constrained by the computational cost to numerically solve the associated equations, with models limited to representing mm-scale or smaller tumors. While the work has proven useful to the study of small tumors and micro-metastases, a biologically-relevant simulation of cm-scale masses as would be typically detected and treated in patients has remained an elusive goal. This study presents a distributed computing (parallelized) implementation of a mixture model of tumor growth to simulate 3D cm-scale vascularized tissue at sub-mm resolution. The numerical solving scheme utilizes a two-stage parallelization framework. The solution is written for GPU computation using the CUDA framework, which handles all Multigrid-related computations. Message Passing Interface (MPI) handles distribution of information across multiple processes, freeing the program from RAM and the processing limitations found on single systems. On each system, Nvidia's CUDA library allows for fast processing of model data using GPU-bound computing on fewer systems. The results show that a combined MPI-CUDA implementation enables the continuum modeling of cm-scale tumors at reasonable computational cost. Further work to calibrate model parameters to particular tumor conditions could enable simulation of patient-specific tumors for clinical application.

Performance Optimization on GPGPU & Multicore CPU Using Roofline Model

The roofline model introduced in this paper to evaluate the best optimized platform for training the neural network that used to recognize handwritten digits under multicore CPU and general-purpose GPU (GPGPU) as hardware environment. The pattern parallel training technique for MNIST dataset is applied. The parallel network training of MNIST using different data layout of multicore CPU and GPGPU is presented. Different bottlenecks have been explained by applying the roofline model. The most suitable platform is selected according to layouts and constrains either for memory or computation bounds. The computational intensity of all rooflines is moved toward right, then the performance is increased. As a result of optimization, and with the diversity of the available data size, core number, operational strength, the most suitable hardware platform is selected.

Multi-GPU acceleration of large-scale density-based topology optimization

This work presents a parallel implementation of density-based topology optimization using distributed GPU computing systems. The use of multiple GPU devices allows us accelerating the computing process and increasing the device memory available for GPU computing. This increment of device memory enables us to address large models that commonly do not fit into one GPU device. The most modern scientific computers incorporate these devices to design energy-efficient, low-cost, and high-computing power systems. However, we should adopt the proper techniques to take advantage of the computational resources of such high-performance many-core computing systems. It is well-known that the bottleneck of density-based topology optimization is the solving of the linear elasticity problem using Finite Element Analysis (FEA) during the topology optimization iterations. We solve the linear system of equations obtained from FEA using a distributed conjugate gradient solver preconditioned by a smooth aggregation-based algebraic multigrid (AMG) using GPU computing with multiple devices. The use of aggregation-based AMG reduces memory requirements and improves the efficiency of the interpolation operation. This fact is rewarding for GPU computing. We evaluate the performance and scalability of the distributed GPU system using structured and unstructured meshes. We also test the performance using different 3D finite elements and relaxing operators. Besides, we evaluate the use of numerical approaches to increase the topology optimization performance. Finally, we present a comparison between the many-core computing instance and one efficient multi-core implementation to highlight the advantages of using GPU computing in large-scale density-based topology optimization problems.

Experiencing GPU path tracing in online courses

In consideration of interdependency between image sensing/recognition (computer vision, CV) and 3D image synthesis (computer graphics, CG) in visual computing, Keio University, Department of Information and Computer Science, reorganized its first course on CV and CG in the undergraduate program into a series of three courses on visual computing in the 2019 academic year. One salient feature of these courses is its newly introduced programming assignment with two specific goals: to experience GPU computing and to understand the path tracing algorithm. The purpose is to help students easily understand the trend in visual computing and vividly envision the future CG area. Specifically, two types of tasks were given to students: material design and analysis of the relationship between sampling and noise. The educational material builds on Google Colaboratory, a cloud-based development environment, to be independent of the students’ hardware. Owing to the judicious design, students can unhesitatingly work on the programming assignment with relatively inexpensive hardware, such as laptop PCs, tablets, or even smartphones, whenever they have a standard off-campus network environment. In the second academic year (2020) of these courses, this type of exercise was especially valuable for students who had to take the courses online because of the COVID-19 pandemic. The effects of the new courses with the educational materials were empirically proven in terms of quantitative and qualitative perspectives.

Towards multiscale modeling of ocean surface turbulent mixing using coupled MPAS-Ocean v6.3 and PALM v5.0

Abstract. A multiscale modeling approach for studying the ocean surface turbulent mixing is explored by coupling an ocean general circulation model (GCM) MPAS-Ocean with the Parallelized Large Eddy Simulation Model (PALM). The coupling approach is similar to the superparameterization approach that has been used to represent the effects of deep convection in atmospheric GCMs. However, the focus of this multiscale modeling approach is on the small-scale turbulent mixing and their interactions with the larger-scale processes in the ocean, so that a more flexible coupling strategy is used. To reduce the computational cost, a customized version of PALM is ported on the general-purpose graphics processing unit (GPU) with OpenACC, achieving 10–16 times overall speedup as compared to running on a single CPU. Even with the GPU-acceleration technique, a superparameterization-like approach to represent the ocean surface turbulent mixing in GCMs using embedded high fidelity and three-dimensional large eddy simulations (LESs) over the global ocean is still computationally intensive and infeasible for long simulations. However, running PALM regionally on selected MPAS-Ocean grid cells is shown to be a promising approach moving forward. The flexible coupling between MPAS-Ocean and PALM allows further exploration of the interactions between the ocean surface turbulent mixing and larger-scale processes, as well as future development and improvement of ocean surface turbulent mixing parameterizations for GCMs.

A Shader-Based Ray Tracing Engine

Recently, ray tracing techniques have been highly adopted to produce high quality images and animations. In this paper, we present our design and implementation of a real-time ray-traced rendering engine. We achieved real-time capability for triangle primitives, based on the ray tracing techniques on GPGPU (general-purpose graphics processing unit) compute shaders. To accelerate the ray tracing engine, we used a set of acceleration techniques, including bounding volume hierarchy, its roped representation, joint up-sampling, and bilateral filtering. Our current implementation shows remarkable speed-ups, with acceptable error values. Experimental results shows 2.5–13.6 times acceleration, and less than 3% error values for the 95% confidence range. Our next step will be enhancing bilateral filter behaviors.

Implementation and Performance Evaluation of RNS Variants of the BFV Homomorphic Encryption Scheme

Homomorphic encryption is an emerging form of encryption that provides the ability to compute on encrypted data without ever decrypting them. Potential applications include aggregating sensitive encrypted data on a cloud environment and computing on the data in the cloud without compromising data privacy. There have been several recent advances resulting in new homomorphic encryption schemes and optimized variants. We implement and evaluate the performance of two optimized variants, namely Bajard-Eynard-Hasan-Zucca (BEHZ) and Halevi-Polyakov-Shoup (HPS), of the most promising homomorphic encryption scheme in CPU and GPU. The most interesting (and also unexpected) result of our performance evaluation is that the HPS variant in practice scales significantly better (typically by 15-30 percent) with increase in multiplicative depth of the computation circuit than BEHZ, implying that the HPS variant will always outperform BEHZ for most practical applications. For the multiplicative depth of 98, our fastest GPU implementation performs homomorphic multiplication in 51 ms for 128-bit security settings, which is faster by two orders of magnitude than prior results and already practical for cloud environments supporting GPU computations. Large multiplicative depths supported by our implementations are required for applications involving deep neural networks, logistic regression learning, and other important machine learning problems.

DYRE: a DYnamic REconfigurable solution to increase GPGPU\u2019s reliability

General-purpose graphics processing units (GPGPUs) are extensively used in high-performance computing. However, it is well known that these devices’ reliability may be limited by the rising of faults at the hardware level. This work introduces a flexible solution to detect and mitigate permanent faults affecting the execution units in these parallel devices. The proposed solution is based on adding some spare modules to perform two in-field operations: detecting and mitigating faults. The solution takes advantage of the regularity of the execution units in the device to avoid significant design changes and reduce the overhead. The proposed solution was evaluated in terms of reliability improvement and area, performance, and power overhead costs. For this purpose, we resorted to a micro-architectural open-source GPGPU model (FlexGripPlus). Experimental results show that the proposed solution can extend the reliability by up to 57%, with overhead costs lower than 2% and 8% in area and power, respectively.

GPGPU Task Scheduling Technique for Reducing the Performance Deviation of Multiple GPGPU Tasks in RPC-Based GPU Virtualization Environments

In remote procedure call (RPC)-based graphic processing unit (GPU) virtualization environments, GPU tasks requested by multiple-user virtual machines (VMs) are delivered to the VM owning the GPU and are processed in a multi-process form. However, because the thread executing the computing on general GPUs cannot arbitrarily stop the task or trigger context switching, GPU monopoly may be prolonged owing to a long-running general-purpose computing on graphics processing unit (GPGPU) task. Furthermore, when scheduling tasks on the GPU, the time for which each user VM uses the GPU is not considered. Thus, in cloud environments that must provide fair use of computing resources, equal use of GPUs between each user VM cannot be guaranteed. We propose a GPGPU task scheduling scheme based on thread division processing that supports GPU use evenly by multiple VMs that process GPGPU tasks in an RPC-based GPU virtualization environment. Our method divides the threads of the GPGPU task into several groups and controls the execution time of each thread group to prevent a specific GPGPU task from a long time monopolizing the GPU. The efficiency of the proposed technique is verified through an experiment in an environment where multiple VMs simultaneously perform GPGPU tasks.

An FPGA-Based LSTM Acceleration Engine for Deep Learning Frameworks

Over the past two decades, Long Short-Term Memory (LSTM) networks have been used to solve problems that require modeling of long sequence because they can selectively remember certain patterns over a long period, thus outperforming traditional feed-forward neural networks and Recurrent Neural Network (RNN) on learning long-term dependencies. However, LSTM is characterized by feedback dependence, which limits the high parallelism of general-purpose processors such as CPU and GPU. Besides, in terms of the energy efficiency of data center applications, the high consumption of GPU and CPU computing cannot be ignored. To deal with the above problems, Field Programmable Gate Array (FPGA) is becoming an ideal alternative. FPGA has the characteristics of low power consumption and low latency, which are helpful for the acceleration and optimization of LSTM and other RNNs. This paper proposes an implementation scheme of the LSTM network acceleration engine based on FPGA and further optimizes the implementation through fixed-point arithmetic, systolic array and lookup table for nonlinear function. On this basis, for easy deployment and application, we integrate the proposed acceleration engine into Caffe, one of the most popular deep learning frameworks. Experimental results show that, compared with CPU and GPU, the FPGA-based acceleration engine can achieve performance improvement of 8.8 and 2.2 times and energy efficiency improvement of 16.9 and 9.6 times, respectively, within Caffe framework.

The Impact of Vertical Densification on Public Lighting in Informal Settlements: Using Virtual Environments as an Evaluation Tool for Policy Making

There are a variety of reasons to support the premise that public lighting is beneficial to urban communities. At the same time, a key challenge for the provision of public lighting in informal settlements is their constant physical transformation. The aim of this paper is to evaluate the application of virtual environments (VEs) in lighting planning and policy making. Despite the fact that VEs offer the opportunity to explore an environment by freely navigating through it, including environments that change over time, this feature is rarely taken into account in decision-making processes. A VE-based analysis tool for informal settlement lighting is presented using a case-study street in the informal settlement of Caracolí in Bogotá as its basis. A household survey, spatial measurements, participants’ observations, and luminous intensity distribution curves constituted the main data set for the VE scenarios. Time-related data on the incremental construction of Caracolí’s informal dwellings was collected by the household survey and that data was then projected into past, present, and future night-time scenarios. The lighting quality of these different scenarios was systematically evaluated via lighting calculation software, revealing a variety of shortcomings caused by the current lighting approach. Based on these findings, an alternative lighting approach was developed and re-examined using lighting calculations. Finally, custom game-engine technology and GPU computing were deployed, which allowed for real-time visualisation of the different lighting scenarios and their lighting quality. This setup therefore enables fast iterative feedback loops for current and future lighting policy scenarios and the resulting lighting design. In the first instance, a VE can illustrate well how current lighting policy results in a significant delay of lighting provision in the early stage of a settlement as well as highlighting the mismatch between lighting technology and the built environment during the vertical densification phases. Second, the VE is able to showcase alternative lighting technologies and policy approaches as well as the resulting lighting effects, enabling a visual comparison of different policy scenarios over several decades. In conclusion it will be argued that the dynamic VE technology appears to be a promising decision-making tool for illustrating potential planning and design shortcomings to policy stakeholders in a manner understandable to the layman.

Super-Resolution Ultrasound Localization Microscopy Through Deep Learning.

Ultrasound localization microscopy has enabled super-resolution vascular imaging through precise localization of individual ultrasound contrast agents (microbubbles) across numerous imaging frames. However, analysis of high-density regions with significant overlaps among the microbubble point spread responses yields high localization errors, constraining the technique to low-concentration conditions. As such, long acquisition times are required to sufficiently cover the vascular bed. In this work, we present a fast and precise method for obtaining super-resolution vascular images from high-density contrast-enhanced ultrasound imaging data. This method, which we term Deep Ultrasound Localization Microscopy (Deep-ULM), exploits modern deep learning strategies and employs a convolutional neural network to perform localization microscopy in dense scenarios, learning the nonlinear image-domain implications of overlapping RF signals originating from such sets of closely spaced microbubbles. Deep-ULM is trained effectively using realistic on-line synthesized data, enabling robust inference in-vivo under a wide variety of imaging conditions. We show that deep learning attains super-resolution with challenging contrast-agent densities, both in-silico as well as in-vivo. Deep-ULM is suitable for real-time applications, resolving about 70 high-resolution patches ( 128×128 pixels) per second on a standard PC. Exploiting GPU computation, this number increases to 1250 patches per second.

Fast Open-Source Toolkit for Water T2 Mapping in the Presence of Fat From Multi-Echo Spin-Echo Acquisitions for Muscle MRI.

Imaging has become a valuable tool in the assessment of neuromuscular diseases, and, specifically, quantitative MR imaging provides robust biomarkers for the monitoring of disease progression. Quantitative evaluation of fat infiltration and quantification of the T2 values of the muscular tissue's water component (wT2) are two of the most essential indicators currently used. As each voxel of the image can contain both water and fat, a two-component model for the estimation of wT2 must be used. In this work, we present a fast method for reconstructing wT2 maps obtained from conventional multi-echo spin-echo (MESE) acquisitions and released as Free Open Source Software. The proposed software is capable of fast reconstruction thanks to extended phase graphs (EPG) simulations and dictionary matching implemented on a general-purpose graphic processing unit. The program can also perform more conventional biexponential least-squares fitting of the data and incorporate information from an external water-fat acquisition to increase the accuracy of the results. The method was applied to the scans of four healthy volunteers and five subjects suffering from facioscapulohumeral muscular dystrophy (FSHD). Conventional multi-slice MESE acquisitions were performed with 17 echoes, and additionally, a 6-echo multi-echo gradient-echo (MEGE) sequence was used for an independent fat fraction calculation. The proposed reconstruction software was applied on the full datasets, and additionally to reduced number of echoes, respectively, to 8, 5, and 3, using EPG and biexponential least-squares fitting, with and without incorporating information from the MEGE acquisition. The incorporation of external fat fraction maps increased the robustness of the fitting with a reduced number of echoes per datasets, whereas with unconstrained fitting, the total of 17 echoes was necessary to retain an independence of wT2 from the level of fat infiltration. In conclusion, the proposed software can successfully be used to calculate wT2 maps from conventional MESE acquisition, allowing the usage of an optimized protocol with similar precision and accuracy as a 17-echo acquisition. As it is freely released to the community, it can be used as a reference for more extensive cohort studies.

DiscretizationNet: A machine-learning based solver for Navier–Stokes equations using finite volume discretization

Over the last few decades, existing Partial Differential Equation (PDE) solvers have demonstrated a tremendous success in solving complex, non-linear PDEs. Although accurate, these PDE solvers are computationally costly. With the advances in Machine Learning (ML) technologies, there has been a significant increase in the research of using ML to solve PDEs. The goal of this work is to develop an ML-based PDE solver, that couples’ important characteristics of existing PDE solvers with ML technologies. The two solver characteristics that have been adopted in this work are: (1) the use of discretization-based schemes to approximate spatio-temporal partial derivatives and (2) the use of iterative algorithms to solve linearized PDEs in their discrete form. In the presence of highly non-linear, coupled PDE solutions, these strategies can be very important in achieving good accuracy, better stability and faster convergence. Our ML-solver, DiscretizationNet, employs a generative CNN-based encoder–decoder model with PDE variables as both input and output features. During training, the discretization schemes are implemented inside the computational graph to enable faster GPU computation of PDE residuals, which are used to update network weights that result into converged solutions. A novel iterative capability is implemented during the network training to improve the stability and convergence of the ML-solver. The ML-Solver is demonstrated to solve the steady, incompressible Navier–Stokes equations in 3-D for several cases such as, lid-driven cavity, flow past a cylinder and conjugate heat transfer.

Study on the thermal dissipation performance of GPU cooling system with nanofluid as coolant

The graphic processing unit is the active component that is an important heat source in electronic devices. The GPU cooling system's thermal dissipation performance with de-ionized water and nanofluid as coolants are presented. In the experiment, the GPU cooling system was embedded with the personal computer's GPU board, which in the overclock operating condition. The Ag nanofluid with 0.015% by volume concentration and de-ionized water is used as coolants flowing in the cooling system, considering the effects of coolant types and heat sink configurations on the thermal resistance. It found that the nanoparticles suspending in the base fluid significantly impact the nanofluid's physical properties and flow features, which results in higher removal capacity and, consequently, lower thermal resistance. The heat sink with a lower flow channel gives lower thermal resistance than those from higher ones. The results from this work have been used to develop a knowledge base that can be used to a design-related thermal cooling system for GPU or other components of electronic devices.