Wojciech Bożejko, Zdziław Hejducki, Mariusz Uchroński et al. · Procedia Computer Science · 2012
We propose the new framework of the distributed tabu search metaheuristic designed to be executed using a multi-GPU cluster, i.e. cluster of nodes equipped with GPU computing units. We propose a hybrid single-walk parallelization of the tabu search, where hybridization consists in examining a number of solutions from a neighborhood concurrently by several GPUs (multi-GPU). The methodology is designed to solve the flexible job shop scheduling problem, diffcult problem of discrete optimization.
Cited by 11Open access
Bingyao Li, Jieming Yin, Youtao Zhang et al. · 2021
In recent years, the ever-growing application complexity and input dataset sizes have driven the popularity of multi-GPU systems as a desirable computing platform for many application domains. While employing multiple GPUs intuitively exposes substantial parallelism for the application acceleration, the delivered performance rarely scales with the number of GPUs. One of the major challenges behind is the address translation efficiency. Many prior works focus on CPUs or single GPU execution scenarios while the address translation in multi-GPU systems receives little attention. In this paper, we conduct a comprehensive investigation of the address translation efficiency in both “single-application-multi-GPU” and “multi-application-multi-GPU” execution paradigms. Based on our observations, we propose a new TLB hierarchy design, called least-TLB, tailored for multi-GPU systems and effectively improves the TLB performance with minimal hardware overheads. Experimental results on 9 single-application workloads and 10 multi-application workloads indicate the proposed least-TLB improves the performances, on average, by 23.5% and 16.3%, respectively.
Cited by 28
Takuya Harada, Tadahiko Muarata · 2017
In this paper, we propose parallel computing techniques for reproducible large-scale social simulations on various computing environments including CPU (Central Processing Unit) or GPU (Graphic Processing Unit). When we use computing resources for large-scale social simulations, the reproducibility of a simulation should be considered. “Reproducibility” means the same trial of a simulation can be repeated. If the same computing resources are available to repeat the trial, it is easy to reproduce the same simulation results. When not all the same computing resources are available, however, it becomes difficult to obtain the same trial since random number generators may become different from the original computation resources. In this study, we employ multi-thread computing on CPU or GPU. We propose two models to run reproducible social simulations on CPU or GPU. One is to parallelize trials (Trial Parallelization). The other is to parallelize agents of a single simulation (Agent Parallelization). These models can be ensured reproducibility even in different computing resources. Our experimental results show that the same computing processes are obtained on CPU or GPU. When we parallelize large-scale social simulation on CPU or GPU, we can accelerate the simulation as a secondary effect.
Cited by 0
Chi Ho Chan, Francis C. M. Lau · 2012
In this paper, the Sum Product Algorithm (SPA) and the Min-Sum Algorithm (MSA) are used for decoding low-density parity-check convolutional codes (LDPC-CCs). The two algorithms have been implemented and run on three different computing environments. The first environment is a single-threading Central Processing Unit (CPU); the second one is the multi-threading CPU based on OpenMP (Open Multi-Processing); and the third one is the multi-threading Graphics Processing Unit (GPU). The error performance of the LDPC-CCs and the simulation time taken under the three specific computing environments and the two decoding algorithms are evaluated and compared. It is found that the different computing environments produce very similar error results. It is also concluded that using the GPU computing platform can reduce the simulation time substantially.
Cited by 2
Xiaoguang Liu · 2019 IEEE International Conference on Signal, Information and Data Processing (ICSIDP) · 2019
The Back-Projection Algorithm(BP) algorithm is an accurate time-domain imaging algorithm, but it needs to reduce calculation time when it is applied to multi-input and multi-output radar imaging due to its large amount of calculation. Using GPU computing, taking advantage of its concurrent processing, can greatly improve the computational efficiency, but have the a problem of communication delay with the acquisition front end and communication overhead between the GPU and the CPU. This paper describes a data transmission system that allows echo data to be directly transmitted into the GPU using RDMA technology through multiple high-speed transceivers and PCIe bus. The echo data loading time is greatly reduced because of the high data transmission bandwidth of multiple transceivers and PCIe. Moreover, multiple GPUs are used for collaborative parallel computing. A ring topology is used to divide the computing tasks, and P2P technology is used for processors communicate directly, which reducing the communication loop between CPU and GPU, basically achieving linear acceleration performance.
Cited by 3
Sungbin Jang, Junhyeok Park, Osang Kwon et al. · 2024
GPU memory virtualization has become essential for efficient programming, memory management, and address space sharing among computing devices in heterogeneous systems. Conventional GPU virtual memory systems use multi-level Radix Page Tables (RPTs) to store virtual-to-physical address mapping in device (GPU) memory. When a TLB miss occurs, a page table walker accesses each level of the page table sequentially to find the desired mapping. These sequential accesses significantly degrade performance, adding pressure to the GPU memory hierarchy. To make matters worse, recent computing systems now support five-level RPTs, further increasing the number of memory accesses required per page table walk.
Cited by 5Open access
Guiming Zhang, Jin Xu · ISPRS International Journal of Geo-Information · 2023
Kernel density estimation (KDE) is a commonly used method for spatial point pattern analysis, but it is computationally demanding when analyzing large datasets. GPU-based parallel computing has been adopted to address such computational challenges. The existing GPU-parallel KDE method, however, utilizes only one GPU for parallel computing. Additionally, it assumes that the input data can be held in GPU memory all at once for computation, which is unrealistic when conducting KDE analysis over large geographic areas at high resolution. This study develops a multi-GPU-parallel and tile-based KDE algorithm to overcome these limitations. It exploits multiple GPUs to speedup complex KDE computation by distributing computation across GPUs, and approaches density estimation with a tile-based strategy to bypass the memory bottleneck. Experiment results show that the parallel KDE algorithm running on multiple GPUs achieves significant speedups over running on a single GPU, and higher speedups are achieved on KDE tasks of a larger problem size. The tile-based strategy renders it feasible to estimate high-resolution density surfaces over large areas even on GPUs with only limited memory. Multi-GPU parallel computing and tile-based density estimation, while incurring very little computational overhead, effectively enable conducting KDE for large-scale spatial point pattern analysis on geospatial big data.
Cited by 13Open access
Tianyu Bai, Spencer Davis, Juanjuan Li et al. · 2014
Lattice based cryptography is attractive for its quantum computing resistance and efficient encryption/decryption process. However, the big data problem has perplexed lattice based cryptographic systems with the slow processing speed. This paper intends to analyze one of the major lattice-based cryptographic systems, Nth-degree truncated polynomial ring (NTRU), and accelerate its execution with Graphic Processing Unit (GPU) for acceptable processing performance. Three strategies, including single GPU with zero copy, single GPU with data transfer, and multi-GPU versions are proposed. GPU computing techniques such as stream and zero copy are applied to overlap the computation and communication for possible speedup. Experimental results have demonstrated the effectiveness of GPU acceleration of NTRU. As the number of involved devices increases, better NTRU performance will be achieved.
Cited by 11
Hermawan Bin Fauzan, Riyanarto Sarno, Ahmad Saikhu · International Journal of Power Electronics and Drive Systems/International Journal of Electrical and Computer Engineering · 2024
With the rapid expansion of process mining implementation in global enterprises distributed across numerous branches, there is a critical requirement to develop an application qualified for real-time operation with fast and precise data integration. To address this challenge, computational parallelism emerges as a feasible solution to accelerate data analytics, with graphical processor unit (GPU) computing currently trending for achieving parallelism acceleration. In this study, we developed a process mining application to optimize parallel and distributed process discovery through a combination of central processing unit (CPU) and GPU computing. The use of this computing combination is leveraged for executing multi-windowing threads within multi-instruction, multiple data (MIMD) in the CPU for streaming distributed event logs, using multi-instruction, single data (MISD) within the CPU to deploy a large footprint pipeline to the GPU, and then utilizing single instruction, multiple data (SIMD) to execute global thread discovery within the GPU. This method significantly accelerates performance in real-time distributed discovery. By reducing branch divergence in SIMD on the global thread GPU parallelism, it outperformed local-thread CPU execution in deterministic discovery, speeding up from 10 to 40 times under specific conditions using a novel min-max flag algorithm implemented within the main steps of the process discovery.
Cited by 0Open access
Олександр Миколайович Хіміч, V. A. Sidoruk, Anton Pavliuk · Cybernetics and Computer Technologies · 2025
This article presents an efficient parallel algorithm for LU factorization of large dense matrices based on a two-dimensional block-cyclic data distribution designed for multi-GPU computing environments. The proposed methodology addresses key challenges of large-scale linear algebra computations, including load balancing, minimizing communication overhead, and maximizing computational locality. By distributing matrix blocks cyclically across a GPU process grid, the algorithm ensures uniform workload distribution even for very large problem sizes, thereby avoiding idle time and reducing synchronization delays. The implementation leverages state-of-the-art GPU computing technologies, including CUDA streams, cuBLAS and cuSolver libraries for local numerical kernels, and NCCL for high-performance collective communications using NVLink interconnects. A look-ahead scheduling strategy and overlapping of communication with computation further increase the degree of parallelism, enabling sustained high utilization of GPU resources throughout the factorization pipeline. A detailed theoretical performance model is developed to analyze speedup, scalability, communication costs, and the impact of block size on total execution time. Numerical experiments conducted on an 8-GPU node with NVIDIA RTX 2080 Ti GPUs demonstrate excellent strong scaling, achieving up to 95 % parallel efficiency for matrices of size up to N = 20,000. The results confirm that the proposed multi-GPU LU factorization approach closely approaches the theoretical performance limits and significantly outperforms traditional CPU-based and hybrid CPU–GPU schemes. The method is highly suitable for large-scale scientific and engineering applications requiring fast and robust solution of linear systems, including computational fluid dynamics, structural mechanics, numerical simulation of physical processes, and machine learning workloads. Future research directions include extensions to sparse matrices with adaptive load balancing, mixed-precision acceleration with error correction, and generalization to other matrix factorizations such as Cholesky, QR, and LDL?. Keywords: LU factorization, multi-GPU systems, block-cyclic data distribution, parallel computing, CUDA, cuBLAS, cuSolver, NCCL, high-performance computing, scalability.
Cited by 0Open access
Sebastian Sbîrnă, Liana-Simona Sbîrnă · 2024
The current work presents an in-depth analysis of several optimizations using GPU parallel computing applied to the Jacobi method for solving Poisson partial differential equations in computational fluid dynamics (CFD). We expand on previous CPU-parallelized Jacobi algorithm research, exploring four GPU-optimized Jacobi method variants: single-threaded, multi-threaded, multi-GPU and a norm-based stopping criterion kernel. These implementations are benchmarked against a multi-threaded CPU baseline. Results indicate that, whereas the single-threaded GPU version is slower than the CPU baseline, multi-threaded GPU versions achieve significant speed gains, especially for larger grid sizes. The multi-GPU version doubles memory bandwidth, enhancing performance for extensive computations, despite overhead for smaller matrices. The norm-stopping criterion kernel offers early convergence for small matrices but at a high overhead cost. Profiling confirms a memory-bound bottleneck, suggesting single-precision and optimized memory access as improvements. Ultimately, multi-threaded GPU kernels substantially outperform the CPU baseline for large-scale CFD problems, establishing GPUs as efficient accelerators for the Jacobi algorithm.
Cited by 0
Roeland Wiersema · ArXiv.org · 2026
Solving large dense linear systems and eigenvalue problems is a core requirement in many areas of scientific computing, but scaling these operations beyond a single GPU remains challenging within modern programming frameworks. While highly optimized multi-GPU solver libraries exist, they are typically difficult to integrate into composable, just-in-time (JIT) compiled Python workflows. JAXMg provides multi-GPU dense linear algebra for JAX, enabling Cholesky-based linear solves and symmetric eigendecompositions for matrices that exceed single-GPU memory limits. By interfacing JAX with NVIDIA's cuSOLVERMg through an XLA Foreign Function Interface, JAXMg exposes distributed GPU solvers as JIT-compatible JAX primitives. This design allows scalable linear algebra to be embedded directly within JAX programs, preserving composability with JAX transformations and enabling multi-GPU execution in end-to-end scientific workflows.
Cited by 0Open access
Roeland Wiersema · arXiv (Cornell University) · 2026
Solving large dense linear systems and eigenvalue problems is a core requirement in many areas of scientific computing, but scaling these operations beyond a single GPU remains challenging within modern programming frameworks. While highly optimized multi-GPU solver libraries exist, they are typically difficult to integrate into composable, just-in-time (JIT) compiled Python workflows. JAXMg provides multi-GPU dense linear algebra for JAX, enabling Cholesky-based linear solves and symmetric eigendecompositions for matrices that exceed single-GPU memory limits. By interfacing JAX with NVIDIA's cuSOLVERMg through an XLA Foreign Function Interface, JAXMg exposes distributed GPU solvers as JIT-compatible JAX primitives. This design allows scalable linear algebra to be embedded directly within JAX programs, preserving composability with JAX transformations and enabling multi-GPU execution in end-to-end scientific workflows.
Cited by 0Open access
Zhixiang Liu, Yu Li, Wei Song · Concurrency and Computation Practice and Experience · 2022
Abstract As an improved method of lattice Boltzmann method (LBM), regularized lattice Boltzmann method (RLBM) has been applied to simulate fluid flow. Nevertheless, the performance of RLBM needs to be considered when simulating actual problems. The rise of multicore platforms, especially the popularity of graphics processor units (GPUs), has provided possible implementation solutions for parallel computing. In this article, an RLBM parallel model on the CPU/GPU heterogeneous platforms is proposed. To solve the problem of possible GPU memory shortage, the CPU controls the startup of the kernel function in the RLBM algorithm and participates in the calculation. Due to the characteristics of the algorithm, the entire flow field is divided into CPU computing areas and GPU computing areas according to the given flow field division rules. OpenMP and CUDA are, respectively, applied to CPU and GPU for parallel computing. The startup of the kernel function takes a short time, and the CPU and GPU can be approximately regarded as performing calculations simultaneously. Since the data exchange between the CPU and GPU has a significant impact on performance, a buffer is set at the boundary of the CPU and GPU to reduce the frequency of data exchange. The buffer size determines the number of program iterations before exchanging data. The performance of the algorithm is measured by MFLUPS, and the algorithm is applied to the 3D lid‐driven cavity flow. The obtained results show that the MFLUPS of the algorithm is 25 times that of CPU MFLUPS, and it adds an increment equivalent to CPU MFLUPS compared with GPU MFLUPS. And the algorithm is extended to a multi‐GPU version, which is also applied to the 3D lid‐driven cavity flow. The results obtained show that the MFLUPS of the multi‐GPU version is 1.8 times that of the single GPU.
Cited by 5
Hongbing Chen, Yulin Huang, Jianyu Yang · 2011
Synthetic Aperture Radar(SAR) echo simulator plays an important role in the development, debugging and performance evaluation of radar system. In this paper, an airborne bistatic SAR echo simulation algorithm based on CUDA architecture is proposed. Then a parallel computing platform with GPUs controlled by CPU threads is presented to achieve the echo simulator. The results show that the speed of the echo simulator based on Multi-GPU is 300 times faster than CPU. Multi-GPU platform has the advantage of computing parallelism, which greatly improves the speed of SAR signal processing.
Cited by 6
Yifan Sun, Trinayan Baruah, Saiful A. Mojumder et al. · arXiv (Cornell University) · 2018
The rapidly growing popularity and scale of data-parallel workloads demand a corresponding increase in raw computational power of GPUs (Graphics Processing Units). As single-GPU systems struggle to satisfy the performance demands, multi-GPU systems have begun to dominate the high-performance computing world. The advent of such systems raises a number of design challenges, including the GPU microarchitecture, multi-GPU interconnect fabrics, runtime libraries and associated programming models. The research community currently lacks a publically available and comprehensive multi-GPU simulation framework and benchmark suite to evaluate multi-GPU system design solutions. In this work, we present MGSim, a cycle-accurate, extensively validated, multi-GPU simulator, based on AMD's Graphics Core Next 3 (GCN3) instruction set architecture. We complement MGSim with MGMark, a suite of multi-GPU workloads that explores multi-GPU collaborative execution patterns. Our simulator is scalable and comes with in-built support for multi-threaded execution to enable fast and efficient simulations. In terms of performance accuracy, MGSim differs $5.5\%$ on average when compared against actual GPU hardware. We also achieve a $3.5\times$ and a $2.5\times$ average speedup in function emulation and architectural simulation with 4 CPU cores, while delivering the same accuracy as the serial simulation. We illustrate the novel simulation capabilities provided by our simulator through a case study exploring programming models based on a unified multi-GPU system (U-MGPU) and a discrete multi-GPU system (D-MGPU) that both utilize unified memory space and cross-GPU memory access. We evaluate the design implications from our case study, suggesting that D-MGPU is an attractive programming model for future multi-GPU systems.
Cited by 4Open access
Sungbo Jung, Dar-Jen Chang, Juw Won Park · ACM SIGAPP Applied Computing Review · 2017
Current microprocessor architecture is moving towards multi-core/multi-threaded systems. This trend has led to a surge of interest in using multi-threaded computing devices, such as the Graphics Processing Unit (GPU), for general purpose computing. We can utilize the GPU in computation as a massive parallel co-processor because the GPU consists of multiple cores. The GPU is also an affordable, attractive, and user-programmable commodity. Nowadays a lot of information has been flooded into the digital domain around the world. Huge volume of data, such as digital libraries, social networking services, e-commerce product data, and reviews, etc., is produced or collected every moment with dramatic growth in size. Although the inverted index is a useful data structure that can be used for full text searches or document retrieval, a large number of documents will require a tremendous amount of time to create the index. The performance of document inversion can be improved by multi-thread or multi-core GPU. Our approach is to implement a linear-time, hash-based, single program multiple data (SPMD), document inversion algorithm on the NVIDIA GPU/CUDA programming platform utilizing the huge computational power of the GPU, to develop high performance solutions for document indexing. Our proposed parallel document inversion system shows 2--3 times faster performance than a sequential system on two different test datasets from PubMed abstract and e-commerce product reviews.
Cited by 0Open access
Wenxiang Zhang, He Jianwei, Chao Sun et al. · 2016
Full waveform inversion is based on gradient or Newton nonlinear optimization algorithm to seek best velocity model to match seismic record. In time domain full waveform inversion, it requires many times full wave-equation modeling and backward modeling. In this paper we use CPU/GPU heterogeneous parallel framework to solve FWI computing problem in multi-Core, multi-GPU cluster. There are three level of parallelization, and the first of level of parallelization is to distribute inversion jobs to each cluster node by shot number, the second level of parallelization take advantage of multi-core computer to call GPU device to do forward modeling and backward calculation. And three level of parallelization use GPU/CUDA to assign each GPU thread block to do modeling and backward calculation.
Cited by 1
Kai Cao, Qizhong Wu, Lingling Wang et al. · 2024
Abstract. The graphics processing units (GPUs) are becoming a compelling acceleration strategy for geoscience numerical model due to their powerful computing performance. In this study, AMD’s heterogeneous compute interface for portability (HIP) was implemented to port the GPU acceleration version of the Piecewise Parabolic Method (PPM) solver (GPU-HADVPPM) from the NVIDIA GPUs to China’ s domestically GPU-like accelerators as GPU-HADVPPM4HIP, and further introduced the multi-level hybrid parallelism scheme to improve the total computational performance of the HIP version of CAMx (CAMx-HIP) model on the China’ s domestically heterogeneous cluster. The experimental results show that the acceleration effect of GPU-HADVPPM on the different GPU accelerator is more obvious when the computing scale is larger, and the maximum speedup of GPU-HADVPPM on the domestic GPU-like accelerator is 28.9 times. The hybrid parallelism with a message passing interface (MPI) and HIP enables achieve up to 17.2 times speedup when configure 32 CPU cores and GPU-like accelerators on the domestic heterogeneous cluster. And the OpenMP technology is introduced to further reduce the computation time of CAMx-HIP model by 1.9 times. More importantly, by comparing the simulation results of GPU-HADVPPM on NVIDIA GPUs and domestic GPU-like accelerators, it is found that the simulation results of GPU-HADVPPM on domestic GPU-like accelerators have less difference than the NVIDIA GPUs, and the reason for this difference may be related to the fact that the NVIDIA GPU sacrifices part of the accuracy for improved computing performance. All in all, the domestic GPU-like accelerators are more accuracy for scientific computing in the field of geoscience numerical models. Furthermore, we also exhibit that the data transfer efficiency between CPU and GPU has an important impact on heterogeneous computing, and point out that optimizing the data transfer efficiency between CPU and GPU is one of the important directions to improve the computing efficiency of geoscience numerical models in heterogeneous clusters in the future.
Cited by 2Open access
Wei Zhao, Le Xu, Rui Li et al. · 2016
The application of the multi-GPU and multi-CPU cooperation to computing large scare electromagnetic object provides a substantial acceleration in speed. In this paper, we show that cooperation of multi-GPU and multi-CPU accelerates hybrid time domain physical optics (TDPO) and time domain equivalent edge current (TDEEC), it runs 150 times faster when a single-thread executes TDPO+TDEEC. The parallel model are pretend to demonstrate the high parallel efficiency solution for computational electromagnetics (CEM).
Cited by 1