Beam-tracking simulations have been extensively utilized in the study of collective beam instabilities in circular accelerators.Traditionally,many simulation codes have relied on central processing unit(CPU)-based met...Beam-tracking simulations have been extensively utilized in the study of collective beam instabilities in circular accelerators.Traditionally,many simulation codes have relied on central processing unit(CPU)-based methods,tracking on a single CPU core,or parallelizing the computation across multiple cores via the message passing interface(MPI).Although these approaches work well for single-bunch tracking,scaling them to multiple bunches significantly increases the computational load,which often necessitates the use of a dedicated multi-CPU cluster.To address this challenge,alternative methods leveraging General-Purpose computing on Graphics Processing Units(GPGPU)have been proposed,enabling tracking studies on a standalone desktop personal computer(PC).However,frequent CPU-GPU interactions,including data transfers and synchronization operations during tracking,can introduce communication overheads,potentially reducing the overall effectiveness of GPU-based computations.In this study,we propose a novel approach that eliminates this overhead by performing the entire tracking simulation process exclusively on the GPU,thereby enabling the simultaneous processing of all bunches and their macro-particles.Specifically,we introduce MBTRACK2-CUDA,a Compute Unified Device Architecture(CUDA)ported version of MBTRACK2,which facilitates efficient tracking of single-and multi-bunch collective effects by leveraging the full GPU-resident computation.展开更多
A computational fluid dynamics(CFD)solver for a GPU/CPU heterogeneous architecture parallel computing platform is developed to simulate incompressible flows on billion-level grid points.To solve the Poisson equation,t...A computational fluid dynamics(CFD)solver for a GPU/CPU heterogeneous architecture parallel computing platform is developed to simulate incompressible flows on billion-level grid points.To solve the Poisson equation,the conjugate gradient method is used as a basic solver,and a Chebyshev method in combination with a Jacobi sub-preconditioner is used as a preconditioner.The developed CFD solver shows good performance on parallel efficiency,which exceeds 90%in the weak-scalability test when the number of grid points allocated to each GPU card is greater than 2083.In the acceleration test,it is found that running a simulation with 10403 grid points on 125 GPU cards accelerates by 203.6x over the same number of CPU cores.The developed solver is then tested in the context of a two-dimensional lid-driven cavity flow and three-dimensional Taylor-Green vortex flow.The results are consistent with previous results in the literature.展开更多
Processing large-scale 3-D gravity data is an important topic in geophysics field. Many existing inversion methods lack the competence of processing massive data and practical application capacity. This study proposes...Processing large-scale 3-D gravity data is an important topic in geophysics field. Many existing inversion methods lack the competence of processing massive data and practical application capacity. This study proposes the application of GPU parallel processing technology to the focusing inversion method, aiming at improving the inversion accuracy while speeding up calculation and reducing the memory consumption, thus obtaining the fast and reliable inversion results for large complex model. In this paper, equivalent storage of geometric trellis is used to calculate the sensitivity matrix, and the inversion is based on GPU parallel computing technology. The parallel computing program that is optimized by reducing data transfer, access restrictions and instruction restrictions as well as latency hiding greatly reduces the memory usage, speeds up the calculation, and makes the fast inversion of large models possible. By comparing and analyzing the computing speed of traditional single thread CPU method and CUDA-based GPU parallel technology, the excellent acceleration performance of GPU parallel computing is verified, which provides ideas for practical application of some theoretical inversion methods restricted by computing speed and computer memory. The model test verifies that the focusing inversion method can overcome the problem of severe skin effect and ambiguity of geological body boundary. Moreover, the increase of the model cells and inversion data can more clearly depict the boundary position of the abnormal body and delineate its specific shape.展开更多
This paper aims to solve large-scale and complex isogeometric topology optimization problems that consumesignificant computational resources. A novel isogeometric topology optimization method with a hybrid parallelstr...This paper aims to solve large-scale and complex isogeometric topology optimization problems that consumesignificant computational resources. A novel isogeometric topology optimization method with a hybrid parallelstrategy of CPU/GPU is proposed, while the hybrid parallel strategies for stiffness matrix assembly, equationsolving, sensitivity analysis, and design variable update are discussed in detail. To ensure the high efficiency ofCPU/GPU computing, a workload balancing strategy is presented for optimally distributing the workload betweenCPU and GPU. To illustrate the advantages of the proposedmethod, three benchmark examples are tested to verifythe hybrid parallel strategy in this paper. The results show that the efficiency of the hybrid method is faster thanserial CPU and parallel GPU, while the speedups can be up to two orders of magnitude.展开更多
Conventional gradient-based full waveform inversion (FWI) is a local optimization, which is highly dependent on the initial model and prone to trapping in local minima. Globally optimal FWI that can overcome this limi...Conventional gradient-based full waveform inversion (FWI) is a local optimization, which is highly dependent on the initial model and prone to trapping in local minima. Globally optimal FWI that can overcome this limitation is particularly attractive, but is currently limited by the huge amount of calculation. In this paper, we propose a globally optimal FWI framework based on GPU parallel computing, which greatly improves the efficiency, and is expected to make globally optimal FWI more widely used. In this framework, we simplify and recombine the model parameters, and optimize the model iteratively. Each iteration contains hundreds of individuals, each individual is independent of the other, and each individual contains forward modeling and cost function calculation. The framework is suitable for a variety of globally optimal algorithms, and we test the framework with particle swarm optimization algorithm for example. Both the synthetic and field examples achieve good results, indicating the effectiveness of the framework. .展开更多
Peridynamics(PD)demonstrates unique advantages in addressing fracture problems,however,its nonlocality and meshfree discretization result in high computational and storage costs.Moreover,in its engineering application...Peridynamics(PD)demonstrates unique advantages in addressing fracture problems,however,its nonlocality and meshfree discretization result in high computational and storage costs.Moreover,in its engineering applications,the computational scale of classical GPU parallel schemes is often limited by the finite graphics memory of GPU devices.In the present study,we develop an efficient particle information management strategy based on the cell-linked list method and on this basis propose a subdomain-based GPU parallel scheme,which exhibits outstanding acceleration performance in specific compute kernels while significantly reducing graphics memory usage.Compared to the classical parallel scheme,the cell-linked list method facilitates efficient management of particle information within subdomains,enabling the proposed parallel scheme to effectively reduce graphics memory usage by optimizing the size and number of subdomains while significantly improving the speed of neighbor search.As demonstrated in PD examples,the proposed parallel scheme enhances the neighbor search efficiency dramatically and achieves a significant speedup relative to serial programs.For instance,without considering the time of data transmission,the proposed scheme achieves a remarkable speedup of nearly 1076.8×in one test case,due to its excellent computational efficiency in the neighbor search.Additionally,for 2D and 3D PD models with tens of millions of particles,the graphics memory usage can be reduced up to 83.6%and 85.9%,respectively.Therefore,this subdomain-based GPU parallel scheme effectively avoids graphics memory shortages while significantly improving the computational efficiency,providing new insights into studying more complex large-scale problems.展开更多
Optical coherence tomography(OCT),particularly Swept-Source OCT,is widely employed in medical diagnostics and industrial inspections owing to its high-resolution imaging capabilities.However,Swept-Source OCT 3D imagin...Optical coherence tomography(OCT),particularly Swept-Source OCT,is widely employed in medical diagnostics and industrial inspections owing to its high-resolution imaging capabilities.However,Swept-Source OCT 3D imaging often suffers from stripe artifacts caused by unstable light sources,system noise,and environmental interference,posing challenges to real-time processing of large-scale datasets.To address this issue,this study introduces a real-time reconstruction system that integrates stripe-artifact suppression and parallel computing using a graphics processing unit.This approach employs a frequency-domain filtering algorithm with adaptive anti-suppression parameters,dynamically adjusted through an image quality evaluation function and optimized using a convolutional neural network for complex frequency-domain feature learning.Additionally,a graphics processing unit integrated 3D reconstruction framework is developed,enhancing data processing throughput and real-time performance via a dual-queue decoupling mechanism.Experimental results demonstrate significant improvements in structural similarity(0.92),peak signal-to-noise ratio(31.62 dB),and stripe suppression ratio(15.73 dB)compared with existing methods.On the RTX 4090 platform,the proposed system achieved an end-to-end delay of 94.36 milliseconds,a frame rate of 10.3 frames per second,and a throughput of 121.5 million voxels per second,effectively suppressing artifacts while preserving image details and enhancing real-time 3D reconstruction performance.展开更多
The present research attempted a Large-Eddy Simulation (LES) of airflow over a steep, three-dimensional isolated hill by using the latest multi-cores multi-CPUs systems. As a result, it was found that 1) turbulence si...The present research attempted a Large-Eddy Simulation (LES) of airflow over a steep, three-dimensional isolated hill by using the latest multi-cores multi-CPUs systems. As a result, it was found that 1) turbulence simulations using approximately 50 million grid points are feasible and 2) the use of this system resulted in the achievement of a high computation speed, which exceeded the speed of parallel computation attained by a single CPU on one of the latest supercomputers. Furthermore, LES was conducted by using the multi-GPUs systems. The results of these simulations revealed the following findings: 1) the multi-GPUs environment which used the NVDIA? Tesla M2090 or the M2075 could simulate turbulence in a model with as many as approximately 50 million grid points. 2) The computation speed achieved by the multi-GPUs environments exceeded that by parallel computation which used four to six CPUs of one of the latest supercomputers.展开更多
Peta-scale high-perfomlance computing systems are increasingly built with heterogeneous CPU and GPU nodes to achieve higher power efficiency and computation throughput. While providing unprecedented capabilities to co...Peta-scale high-perfomlance computing systems are increasingly built with heterogeneous CPU and GPU nodes to achieve higher power efficiency and computation throughput. While providing unprecedented capabilities to conduct computational experiments of historic significance, these systems are presently difficult to program. The users, who are domain experts rather than computer experts, prefer to use programming models closer to their domains （e.g., physics and biology） rather than MPI and OpenME This has led the development of domain-specific programming that provides domain-specific programming interfaces but abstracts away some performance-critical architecture details. Based on experience in designing large-scale computing systems, a hybrid programming framework for scientific computing on heterogeneous architectures is proposed in this work. Its design philosophy is to provide a collaborative mechanism for domain experts and computer experts so that both domain-specific knowledge and performance-critical architecture details can be adequately exploited. Two real-world scientific applications have been evaluated on TH-IA, a peta-scale CPU-GPU heterogeneous system that is currently the 5th fastest supercomputer in the world. The experimental results show that the proposed framework is well suited for developing large-scale scientific computing applications on peta-scale heterogeneous CPU/GPU systems.展开更多
Digital data have become a torrent engulfing every area of business, science and engineering disciplines, gushing into every economy, every organization and every user of digital technology. In the age of big data, de...Digital data have become a torrent engulfing every area of business, science and engineering disciplines, gushing into every economy, every organization and every user of digital technology. In the age of big data, deriving values and insights from big data using rich analytics becomes important for achieving competitiveness, success and leadership in every field. The Internet of Things (IoT) is causing the number and types of products to emit data at an unprecedented rate. Heterogeneity, scale, timeliness, complexity, and privacy problems with large data impede progress at all phases of the pipeline that can create value from data issues. With the push of such massive data, we are entering a new era of computing driven by novel and ground breaking research innovation on elastic parallelism, partitioning and scalability. Designing a scalable system for analysing, processing and mining huge real world datasets has become one of the challenging problems facing both systems researchers and data management researchers. In this paper, we will give an overview of computing infrastructure for IoT data processing, focusing on architectural and major challenges of massive data. We will briefly discuss about emerging computing infrastructure and technologies that are promising for improving massive data management.展开更多
This paper presents a comprehensive exploration into the integration of Internet of Things(IoT),big data analysis,cloud computing,and Artificial Intelligence(AI),which has led to an unprecedented era of connectivity.W...This paper presents a comprehensive exploration into the integration of Internet of Things(IoT),big data analysis,cloud computing,and Artificial Intelligence(AI),which has led to an unprecedented era of connectivity.We delve into the emerging trend of machine learning on embedded devices,enabling tasks in resource-limited environ-ments.However,the widespread adoption of machine learning raises significant privacy concerns,necessitating the development of privacy-preserving techniques.One such technique,secure multi-party computation(MPC),allows collaborative computations without exposing private inputs.Despite its potential,complex protocols and communication interactions hinder performance,especially on resource-constrained devices.Efforts to enhance efficiency have been made,but scalability remains a challenge.Given the success of GPUs in deep learning,lever-aging embedded GPUs,such as those offered by NVIDIA,emerges as a promising solution.Therefore,we propose an Embedded GPU-based Secure Two-party Computation(EG-STC)framework for Artificial Intelligence(AI)systems.To the best of our knowledge,this work represents the first endeavor to fully implement machine learning model training based on secure two-party computing on the Embedded GPU platform.Our experimental results demonstrate the effectiveness of EG-STC.On an embedded GPU with a power draw of 5 W,our implementation achieved a secure two-party matrix multiplication throughput of 5881.5 kilo-operations per millisecond(kops/ms),with an energy efficiency ratio of 1176.3 kops/ms/W.Furthermore,leveraging our EG-STC framework,we achieved an overall time acceleration ratio of 5–6 times compared to solutions running on server-grade CPUs.Our solution also exhibited a reduced runtime,requiring only 60%to 70%of the runtime of previously best-known methods on the same platform.In summary,our research contributes to the advancement of secure and efficient machine learning implementations on resource-constrained embedded devices,paving the way for broader adoption of AI technologies in various applications.展开更多
基金supported by the National Research Foundation of Korea(NRF)funded by the Ministry of Science and ICT(MSIT)(No.RS-2022-00143178)the Ministry of Education(MOE)(Nos.2022R1A6A3A13053896 and 2022R1F1A1074616),Republic of Korea.
文摘Beam-tracking simulations have been extensively utilized in the study of collective beam instabilities in circular accelerators.Traditionally,many simulation codes have relied on central processing unit(CPU)-based methods,tracking on a single CPU core,or parallelizing the computation across multiple cores via the message passing interface(MPI).Although these approaches work well for single-bunch tracking,scaling them to multiple bunches significantly increases the computational load,which often necessitates the use of a dedicated multi-CPU cluster.To address this challenge,alternative methods leveraging General-Purpose computing on Graphics Processing Units(GPGPU)have been proposed,enabling tracking studies on a standalone desktop personal computer(PC).However,frequent CPU-GPU interactions,including data transfers and synchronization operations during tracking,can introduce communication overheads,potentially reducing the overall effectiveness of GPU-based computations.In this study,we propose a novel approach that eliminates this overhead by performing the entire tracking simulation process exclusively on the GPU,thereby enabling the simultaneous processing of all bunches and their macro-particles.Specifically,we introduce MBTRACK2-CUDA,a Compute Unified Device Architecture(CUDA)ported version of MBTRACK2,which facilitates efficient tracking of single-and multi-bunch collective effects by leveraging the full GPU-resident computation.
基金supported by the National Natural Science Foundation of China (NSFC)Basic Science Center Program for Multiscale Problems in Nonlinear Mechanics’(Grant No. 11988102)NSFC project (Grant No. 11972038)
文摘A computational fluid dynamics(CFD)solver for a GPU/CPU heterogeneous architecture parallel computing platform is developed to simulate incompressible flows on billion-level grid points.To solve the Poisson equation,the conjugate gradient method is used as a basic solver,and a Chebyshev method in combination with a Jacobi sub-preconditioner is used as a preconditioner.The developed CFD solver shows good performance on parallel efficiency,which exceeds 90%in the weak-scalability test when the number of grid points allocated to each GPU card is greater than 2083.In the acceleration test,it is found that running a simulation with 10403 grid points on 125 GPU cards accelerates by 203.6x over the same number of CPU cores.The developed solver is then tested in the context of a two-dimensional lid-driven cavity flow and three-dimensional Taylor-Green vortex flow.The results are consistent with previous results in the literature.
基金Supported by the National Science and Technology Major Project of the Ministry of Science and Technology of China （No. 2013ZX06002001- 007）, the National Key Scientific Instrument and Equipment Development Projects, China （No. 2012YQ180118） and the National Natural Science Foundation of China （Nos. 11275110, 11075091 and 11105081）.
基金Supported by Project of National Natural Science Foundation(No.41874134)
文摘Processing large-scale 3-D gravity data is an important topic in geophysics field. Many existing inversion methods lack the competence of processing massive data and practical application capacity. This study proposes the application of GPU parallel processing technology to the focusing inversion method, aiming at improving the inversion accuracy while speeding up calculation and reducing the memory consumption, thus obtaining the fast and reliable inversion results for large complex model. In this paper, equivalent storage of geometric trellis is used to calculate the sensitivity matrix, and the inversion is based on GPU parallel computing technology. The parallel computing program that is optimized by reducing data transfer, access restrictions and instruction restrictions as well as latency hiding greatly reduces the memory usage, speeds up the calculation, and makes the fast inversion of large models possible. By comparing and analyzing the computing speed of traditional single thread CPU method and CUDA-based GPU parallel technology, the excellent acceleration performance of GPU parallel computing is verified, which provides ideas for practical application of some theoretical inversion methods restricted by computing speed and computer memory. The model test verifies that the focusing inversion method can overcome the problem of severe skin effect and ambiguity of geological body boundary. Moreover, the increase of the model cells and inversion data can more clearly depict the boundary position of the abnormal body and delineate its specific shape.
基金the National Key R&D Program of China(2020YFB1708300)the National Natural Science Foundation of China(52005192)the Project of Ministry of Industry and Information Technology(TC210804R-3).
文摘This paper aims to solve large-scale and complex isogeometric topology optimization problems that consumesignificant computational resources. A novel isogeometric topology optimization method with a hybrid parallelstrategy of CPU/GPU is proposed, while the hybrid parallel strategies for stiffness matrix assembly, equationsolving, sensitivity analysis, and design variable update are discussed in detail. To ensure the high efficiency ofCPU/GPU computing, a workload balancing strategy is presented for optimally distributing the workload betweenCPU and GPU. To illustrate the advantages of the proposedmethod, three benchmark examples are tested to verifythe hybrid parallel strategy in this paper. The results show that the efficiency of the hybrid method is faster thanserial CPU and parallel GPU, while the speedups can be up to two orders of magnitude.
文摘Conventional gradient-based full waveform inversion (FWI) is a local optimization, which is highly dependent on the initial model and prone to trapping in local minima. Globally optimal FWI that can overcome this limitation is particularly attractive, but is currently limited by the huge amount of calculation. In this paper, we propose a globally optimal FWI framework based on GPU parallel computing, which greatly improves the efficiency, and is expected to make globally optimal FWI more widely used. In this framework, we simplify and recombine the model parameters, and optimize the model iteratively. Each iteration contains hundreds of individuals, each individual is independent of the other, and each individual contains forward modeling and cost function calculation. The framework is suitable for a variety of globally optimal algorithms, and we test the framework with particle swarm optimization algorithm for example. Both the synthetic and field examples achieve good results, indicating the effectiveness of the framework. .
基金Jun Li was supported by National Natural Science Foundation of China(No.:U2441215)Lisheng Liu and Xin Lai were supported by National Natural Science Foundation of China(No.:52494933).
文摘Peridynamics(PD)demonstrates unique advantages in addressing fracture problems,however,its nonlocality and meshfree discretization result in high computational and storage costs.Moreover,in its engineering applications,the computational scale of classical GPU parallel schemes is often limited by the finite graphics memory of GPU devices.In the present study,we develop an efficient particle information management strategy based on the cell-linked list method and on this basis propose a subdomain-based GPU parallel scheme,which exhibits outstanding acceleration performance in specific compute kernels while significantly reducing graphics memory usage.Compared to the classical parallel scheme,the cell-linked list method facilitates efficient management of particle information within subdomains,enabling the proposed parallel scheme to effectively reduce graphics memory usage by optimizing the size and number of subdomains while significantly improving the speed of neighbor search.As demonstrated in PD examples,the proposed parallel scheme enhances the neighbor search efficiency dramatically and achieves a significant speedup relative to serial programs.For instance,without considering the time of data transmission,the proposed scheme achieves a remarkable speedup of nearly 1076.8×in one test case,due to its excellent computational efficiency in the neighbor search.Additionally,for 2D and 3D PD models with tens of millions of particles,the graphics memory usage can be reduced up to 83.6%and 85.9%,respectively.Therefore,this subdomain-based GPU parallel scheme effectively avoids graphics memory shortages while significantly improving the computational efficiency,providing new insights into studying more complex large-scale problems.
文摘Optical coherence tomography(OCT),particularly Swept-Source OCT,is widely employed in medical diagnostics and industrial inspections owing to its high-resolution imaging capabilities.However,Swept-Source OCT 3D imaging often suffers from stripe artifacts caused by unstable light sources,system noise,and environmental interference,posing challenges to real-time processing of large-scale datasets.To address this issue,this study introduces a real-time reconstruction system that integrates stripe-artifact suppression and parallel computing using a graphics processing unit.This approach employs a frequency-domain filtering algorithm with adaptive anti-suppression parameters,dynamically adjusted through an image quality evaluation function and optimized using a convolutional neural network for complex frequency-domain feature learning.Additionally,a graphics processing unit integrated 3D reconstruction framework is developed,enhancing data processing throughput and real-time performance via a dual-queue decoupling mechanism.Experimental results demonstrate significant improvements in structural similarity(0.92),peak signal-to-noise ratio(31.62 dB),and stripe suppression ratio(15.73 dB)compared with existing methods.On the RTX 4090 platform,the proposed system achieved an end-to-end delay of 94.36 milliseconds,a frame rate of 10.3 frames per second,and a throughput of 121.5 million voxels per second,effectively suppressing artifacts while preserving image details and enhancing real-time 3D reconstruction performance.
文摘The present research attempted a Large-Eddy Simulation (LES) of airflow over a steep, three-dimensional isolated hill by using the latest multi-cores multi-CPUs systems. As a result, it was found that 1) turbulence simulations using approximately 50 million grid points are feasible and 2) the use of this system resulted in the achievement of a high computation speed, which exceeded the speed of parallel computation attained by a single CPU on one of the latest supercomputers. Furthermore, LES was conducted by using the multi-GPUs systems. The results of these simulations revealed the following findings: 1) the multi-GPUs environment which used the NVDIA? Tesla M2090 or the M2075 could simulate turbulence in a model with as many as approximately 50 million grid points. 2) The computation speed achieved by the multi-GPUs environments exceeded that by parallel computation which used four to six CPUs of one of the latest supercomputers.
基金Project(61170049) supported by the National Natural Science Foundation of ChinaProject(2012AA010903) supported by the National High Technology Research and Development Program of China
文摘Peta-scale high-perfomlance computing systems are increasingly built with heterogeneous CPU and GPU nodes to achieve higher power efficiency and computation throughput. While providing unprecedented capabilities to conduct computational experiments of historic significance, these systems are presently difficult to program. The users, who are domain experts rather than computer experts, prefer to use programming models closer to their domains （e.g., physics and biology） rather than MPI and OpenME This has led the development of domain-specific programming that provides domain-specific programming interfaces but abstracts away some performance-critical architecture details. Based on experience in designing large-scale computing systems, a hybrid programming framework for scientific computing on heterogeneous architectures is proposed in this work. Its design philosophy is to provide a collaborative mechanism for domain experts and computer experts so that both domain-specific knowledge and performance-critical architecture details can be adequately exploited. Two real-world scientific applications have been evaluated on TH-IA, a peta-scale CPU-GPU heterogeneous system that is currently the 5th fastest supercomputer in the world. The experimental results show that the proposed framework is well suited for developing large-scale scientific computing applications on peta-scale heterogeneous CPU/GPU systems.
文摘Digital data have become a torrent engulfing every area of business, science and engineering disciplines, gushing into every economy, every organization and every user of digital technology. In the age of big data, deriving values and insights from big data using rich analytics becomes important for achieving competitiveness, success and leadership in every field. The Internet of Things (IoT) is causing the number and types of products to emit data at an unprecedented rate. Heterogeneity, scale, timeliness, complexity, and privacy problems with large data impede progress at all phases of the pipeline that can create value from data issues. With the push of such massive data, we are entering a new era of computing driven by novel and ground breaking research innovation on elastic parallelism, partitioning and scalability. Designing a scalable system for analysing, processing and mining huge real world datasets has become one of the challenging problems facing both systems researchers and data management researchers. In this paper, we will give an overview of computing infrastructure for IoT data processing, focusing on architectural and major challenges of massive data. We will briefly discuss about emerging computing infrastructure and technologies that are promising for improving massive data management.
基金supported in part by Major Science and Technology Demonstration Project of Jiangsu Provincial Key R&D Program under Grant No.BE2023025in part by the National Natural Science Foundation of China under Grant No.62302238+2 种基金in part by the Natural Science Foundation of Jiangsu Province under Grant No.BK20220388in part by the Natural Science Research Project of Colleges and Universities in Jiangsu Province under Grant No.22KJB520004in part by the China Postdoctoral Science Foundation under Grant No.2022M711689.
文摘This paper presents a comprehensive exploration into the integration of Internet of Things(IoT),big data analysis,cloud computing,and Artificial Intelligence(AI),which has led to an unprecedented era of connectivity.We delve into the emerging trend of machine learning on embedded devices,enabling tasks in resource-limited environ-ments.However,the widespread adoption of machine learning raises significant privacy concerns,necessitating the development of privacy-preserving techniques.One such technique,secure multi-party computation(MPC),allows collaborative computations without exposing private inputs.Despite its potential,complex protocols and communication interactions hinder performance,especially on resource-constrained devices.Efforts to enhance efficiency have been made,but scalability remains a challenge.Given the success of GPUs in deep learning,lever-aging embedded GPUs,such as those offered by NVIDIA,emerges as a promising solution.Therefore,we propose an Embedded GPU-based Secure Two-party Computation(EG-STC)framework for Artificial Intelligence(AI)systems.To the best of our knowledge,this work represents the first endeavor to fully implement machine learning model training based on secure two-party computing on the Embedded GPU platform.Our experimental results demonstrate the effectiveness of EG-STC.On an embedded GPU with a power draw of 5 W,our implementation achieved a secure two-party matrix multiplication throughput of 5881.5 kilo-operations per millisecond(kops/ms),with an energy efficiency ratio of 1176.3 kops/ms/W.Furthermore,leveraging our EG-STC framework,we achieved an overall time acceleration ratio of 5–6 times compared to solutions running on server-grade CPUs.Our solution also exhibited a reduced runtime,requiring only 60%to 70%of the runtime of previously best-known methods on the same platform.In summary,our research contributes to the advancement of secure and efficient machine learning implementations on resource-constrained embedded devices,paving the way for broader adoption of AI technologies in various applications.
