• Journal of the Korea Society of Computer and Information
• /
• 제24권4호
• /
• pp.1-9
• /
• 2019

LRR(Loose Round Robin) warp scheduling policy for GPU architecture results in high warp-level parallelism and balanced loads across multiple warps. However, traditional LRR policy makes multiple warps execute long latency operations at the same time. In cases that no more warps to be issued under long latency, the throughput of GPUs may be degraded significantly. In this paper, we propose a new warp scheduling policy which utilizes latency hiding, leading to more utilized memory resources in high performance GPUs. The proposed warp scheduler prioritizes memory instruction based on GTO(Greedy Then Oldest) policy in order to provide reduced memory stalls. When no warps can execute memory instruction any more, the warp scheduler selects a warp for computation instruction by round robin manner. Furthermore, our proposed technique achieves high performance by using additional information about recently committed warps. According to our experimental results, our proposed technique improves GPU performance by 12.7% and 5.6% over LRR and GTO on average, respectively.

• Journal of Information Processing Systems
• /
• 제17권6호
• /
• pp.1157-1169
• /
• 2021

Modern graphics processor unit (GPU) architectures offer significant hardware resource enhancements for parallel computing. However, without software optimization, GPUs continuously exhibit hardware resource underutilization. In this paper, we indicate the need to alter different warp scheduler schemes during different kernel execution periods to improve resource utilization. Existing warp schedulers cannot be aware of the kernel progress to provide an effective scheduling policy. In addition, we identified the potential for improving resource utilization for multiple-warp-scheduler GPUs by sharing stalling warps with selected warp schedulers. To address the efficiency issue of the present GPU, we coordinated the kernel-aware warp scheduler and warp sharing mechanism (KAWS). The proposed warp scheduler acknowledges the execution progress of the running kernel to adapt to a more effective scheduling policy when the kernel progress attains a point of resource underutilization. Meanwhile, the warp-sharing mechanism distributes stalling warps to different warp schedulers wherein the execution pipeline unit is ready. Our design achieves performance that is on an average higher than that of the traditional warp scheduler by 7.97% and employs marginal additional hardware overhead.

• Journal of Internet Computing and Services
• /
• 제14권6호
• /
• pp.41-47
• /
• 2013

GP-GPUs are general purposed GPUs for numerical computation based on multiple threads which are originally for graphic processing. GP-GPUs provide cache memory in a form of shared memory which user programs can access directly, unlikely typical cache memory. In this research, we implemented the parallel block LU decomposition program to utilize cache memory in GP-GPUs. The parallel blocked LU decomposition program designed with Nvidia CUDA C run 7~8 times faster than nun-blocked LU decomposition program in the same GP-GPU computation environment.

• Journal of Korea Multimedia Society
• /
• 제14권11호
• /
• pp.1458-1466
• /
• 2011

Recently the research on GPGPU has been carried out actively as the performance of GPUs has been increased rapidly. In this paper, we propose the system architecture by benchmarking the existing supercomputer architecture for a cost-effective system using GPUs in low-cost graphics devices and implement a GPU cluster system with eight GPUs. We also make the software development environment that is suitable for the GPU cluster system and use it for the performance evaluation by implementing the n-body problem. According to its result, we found that it is efficient to use multiple GPUs when the problem size is large due to its communication cost. In addition, we could calculate up to eight million celestial bodies by applying the method of calculating block by block to mitigate the problem size constraint due to the limited resource in GPUs.

• IEIE Transactions on Smart Processing and Computing
• /
• 제4권1호
• /
• pp.10-15
• /
• 2015

This paper proposes the parallel implementation of a self-similarity based image SR (super-resolution) algorithm using OpenCL. The SR algorithm requires tremendous computations to search for a similar patch. This becomes a bottleneck for the real-time conversion from a FHD image to UHD. Therefore, it is imperative to accelerate the processing speed of SR algorithms. For parallelization, the SR process is divided into several kernels, and memory optimization is performed. In addition, two GPUs are used for further acceleration. The experimental results shows that a GPGPU implementation can speed up over 140 times compared to a single-core CPU. Furthermore, it was confirmed experimentally that utilizing two GPUs can speed up the execution time proportionally, up to 277 times.

• IEIE Transactions on Smart Processing and Computing
• /
• 제5권4호
• /
• pp.294-301
• /
• 2016

In this paper, we present methods to efficiently parallelize iterative 3D image reconstruction by exploiting trigeneous devices (three different types of device) at the same time: a CPU, an integrated GPU, and a discrete GPU. We first present a technique that exploits single instruction multiple data (SIMD) architectures in GPUs. Then, we propose a performance estimation model, based on which we can easily find the optimal data partitioning on trigeneous devices. We found that the performance significantly varies by up to 6.23 times, depending on how SIMD units in GPUs are accessed. Then, by using trigeneous devices and the proposed estimation models, we achieve optimal partitioning and throughput, which corresponds to a 9.4% further improvement, compared to discrete GPU-only execution.

• Journal of information and communication convergence engineering
• /
• 제20권2호
• /
• pp.96-102
• /
• 2022

We aim to build predictive models for Airbnb's prices using a GPU-accelerated RAPIDS in a big data cluster. The Airbnb Listings datasets are used for the predictive analysis. Several machine-learning algorithms have been adopted to build models that predict the price of Airbnb listings. We compare the results of traditional and big data approaches to machine learning for price prediction and discuss the performance of the models. We built big data models using Databricks Spark Cluster, a distributed parallel computing system. Furthermore, we implemented models using multiple GPUs using RAPIDS in the spark cluster. The model was developed using the XGBoost algorithm, whereas other models were developed using traditional central processing unit (CPU)-based algorithms. This study compared all models in terms of accuracy metrics and computing time. We observed that the XGBoost model with RAPIDS using GPUs had the highest accuracy and computing time.

• Smart Media Journal
• /
• 제8권2호
• /
• pp.9-15
• /
• 2019

Since the computational complexity for hologram generation increases exponentially with respect to the size of the point cloud, parallel processing using CUDA and/or OpenCL library based on multiple GPUs has recently become popular. The CUDA kernel for parallelization needs to consist of threads, blocks, and grids properly in accordance with the number of cores and the memory size in the GPU. In addition, in case of multiple GPU environments, the distribution in grid-by-grid, in block-by-block, or in thread-by-thread is needed according to the number of GPUs. In order to evaluate the performance of CGH generation, we compared the computational speed in CPU, in single GPU, and in multi-GPU environments by gradually increasing the number of points in a point cloud from 10 to 1,000,000. We also present a memory structure design and a calculation method required in the CUDA-based parallel processing to accelerate the CGH (Computer Generated Hologram) generation operation in multiple GPU environments.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• 제66권6호
• /
• pp.955-961
• /
• 2017

In this paper, we propose a parallel approximate string matching algorithm with k-mismatches for multiple fixed-length patterns (PMASM) in DNA sequences. PMASM is developed from parallel single pattern approximate string matching algorithms to effectively calculate the Hamming distances for multiple patterns with a fixed-length. In the preprocessing phase of PMASM, all target patterns are binary encoded and stored into a look-up memory. With each input character from the input string, the Hamming distances between a substring and all patterns can be updated at the same time based on the binary encoding information in the look-up memory. Moreover, PMASM adopts graphics processing units (GPUs) to process the data computations in parallel. This paper presents three kinds of PMASM implementation methods in GPUs: thread PMASM, block-thread PMASM, and shared-mem PMASM methods. The shared-mem PMASM method gives an example to effectively make use of the GPU parallel capacity. Moreover, it also exploits special features of the CUDA (Compute Unified Device Architecture) memory structure to optimize the performance. In the experiments with DNA sequences, the proposed PMASM on GPU is 385, 77, and 64 times faster than the traditional naive algorithm, the shift-add algorithm and the single thread PMASM implementation on CPU. With the same NVIDIA GPU model, the performance of the proposed approach is enhanced up to 44% and 21%, compared with the naive, and the shift-add algorithms.

• Journal of KIISE
• /
• 제42권4호
• /
• pp.522-529
• /
• 2015

Ever since the R-tree was proposed to index multi-dimensional data, many efforts have been made to improve its query performances. One common trend to improve query performance is to parallelize query processing with the use of multi-core architectures. To this end, a GPU-base R-tree has been recently proposed. However, even though a GPU-based R-tree can exhibit an improvement in query performance, it is limited in its ability to handle large volumes of data because GPUs have limited physical memory. To address this problem, we propose MGR-tree (Multi-GPU R-tree), which can manage large volumes of data by dividing nodes into multiple GPUs. Our experiments show that MGR-tree is up to 9.1 times faster than a sequential search on a GPU and up to 1.6 times faster than a conventional GPU-based R-tree.