General-purpose Computing On The Graphics Processing Unit Research Articles

GPUart - An application-based limited preemptive GPU real-time scheduler for embedded systems

Emerging technologies like autonomous driving entail computational intense software solutions. More and more companies accelerate their embedded applications by General Purpose Computing on the Graphics Processing Unit (GPGPU), in order to overcome those computational demands. Unfortunately, Graphics Processing Units (GPUs) severely lack real-time capability, for example controllable preemption support, which limits their applicability in the embedded domain. We therefore present GPUart, a framework for GPU real-time scheduling. GPUart focuses on embedded systems and requires neither hardware nor driver stack extensions. We propose a software-only approach for preemption, based on the fixed preemption point strategy. In contrast to prior work, GPUart enables preemption inside a thread block by adding fixed preemption points. We further propose a portable high-level resource management concept to enable gang scheduling on GPUs. GPUart can schedule GPU workload either under the Gang-Earliest Deadline First (EDF) or Gang-Fixed Task Priority (FTP) policy. A case-study on Nvidia Tegra X1, using real-world engine management applications from Audi AG and Continental Automotive GmbH, shows that only up to 0.28% additional global memory is required to enable interruptible thread blocks. GPUart reduces the worst observed response times by a factor of up to 221, leading to response times without deadline misses.

Brain big data processing with massively parallel computing technology: challenges and opportunities

SummaryBrain data processing has been embracing the big data era driven by the rapid advances of neuroscience as well as the experimental techniques for recording neuronal activities. Processing of massive brain data has become a constant in neuroscience research and practice, which is vital in revealing the hidden information to better understand the brain functions and malfunctions. Brain data are routinely non‐linear and non‐stationary in nature, and existing algorithms and approaches to neural data processing are generally complicated in order to characterize the non‐linearity and non‐stationarity. Brain big data processing has pressing needs for appropriate computing technologies to address three grand challenges: (1) efficiency, (2) scalability and (3) reliability. Recent advances of computing technologies are making non‐linear methods viable in sophisticated applications of massive brain data processing. General‐purpose Computing on the Graphics Processing Unit (GPGPU) technology fosters an ideal environment for this purpose, which benefits from the tremendous computing power of modern graphics processing units in massively parallel architecture that is frequently an order of magnitude larger than the modern multi‐core CPUs. This article first recaps significant speed‐ups of existing algorithms aided by GPGPU in neuroimaging and processing electroencephalogram (EEG), functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG) and etc. The article then demonstrates a series of successful approaches to processing EEG data in various dimensions and scales in a massively parallel manner: (1) decomposition: a massively parallel Ensemble Local Mean Decomposition (ELMD) algorithm aided by GPGPU can decompose EEG series, which forms the basis of further time‐frequency transformation, in real‐time without sacrificing the precision of processing; (2) synchronization measurement: a parallelized Nonlinear Interdependence (NLI) method for global synchronization measurement of multivariate EEG with speed‐up of more than 1000 times, and it was successful in localization of epileptic focus; and (3) dimensionality reduction: a large‐scale Parallel Factor Analysis which excels in run‐time performance and scales far better by hundreds of times than conventional approach does, and it supports fast factorization of EEG with more than 1000 channels. Through these practices, the massively parallel computing technology manifests great potentials in addressing the grand challenges of brain big data processing. Copyright © 2016 John Wiley & Sons, Ltd.

G-IK-SVD: parallel IK-SVD on GPUs for sparse representation of spatial big data

Sparse representation is a building block for many image processing applications such as compression, denoising, fusion and so on. In the era of "Big data", the current spare representation methods generally do not meet the demand of time-efficiently processing the large image dataset. Aiming at this problem, this paper employed the contemporary general-purpose computing on the graphics processing unit (GPGPU) to extend a sparse representation method for big image datasets, IK-SVD, namely G-IK-SVD. The GPU-aided IK-SVD parallelized IK-SVD with three GPU optimization methods: (1) a batch-OMP algorithm based on GPU-aided Cholesky decomposition algorithm, (2) a GPU sparse matrix operation optimization method and (3) a hybrid parallel scheme. The experimental results indicate that (1) the GPU-aided batch-OMP algorithm shows speedups of up to 30 times than the sparse coding part of IK-SVD, (2) the optimized sparse matrix operations improve the whole procedure of IK-SVD up to 15 times,(3) the proposed parallel scheme can further accelerate the procedure of sparsely representing one large image dataset up to 24 times, and (4) G-IK-SVD can gain the same quality of dictionary learning as IK-SVD.

Fast and Scalable Multi-Way Analysis of Massive Neural Data

Analysis of neural data with multiple modes and high density has recently become a trend with the advances in neuroscience research and practices. There exists a pressing need for an approach to accurately and uniquely capture the features without loss or destruction of the interactions amongst the modes (typically) of space, time, and frequency. Moreover, the approach must be able to quickly analyze the neural data of exponentially growing scales and sizes, in tens or even hundreds of channels, so that timely conclusions and decisions may be made. A salient approach to multi-way data analysis is the parallel factor analysis (PARAFAC) that manifests its effectiveness in the decomposition of the electroencephalography (EEG). However, the conventional PARAFAC is only suited for offline data analysis due to the high complexity, which computes to be $O(n^{2})$ with the increasing data size. In this study, a large-scale PARAFAC method has been developed, which is supported by general-purpose computing on the graphics processing unit (GPGPU). Comparing to the PARAFAC running on conventional CPU-based platform, the new approach dramatically excels by ${>}360$ times in run-time performance, and effectively scales by ${>}400$ times in all dimensions. Moreover, the proposed approach forms the basis of a model for the analysis of electrocochleography (ECoG) recordings obtained from epilepsy patients, which proves to be effective in the epilepsy state detection. The time evolutions of the proposed model are well correlated with the clinical observations. Moreover, the frequency signature is stable and high in the ictal phase. Furthermore, the spatial signature explicitly identifies the propagation of neural activities among various brain regions. The model supports real-time analysis of ECoG in ${>}1{,}000$ channels on an inexpensive and available cyber-infrastructure.

Parallel Processing of Dynamic Continuous Queries over Streaming Data Flows

More and more real-time applications need to handle dynamic continuous queries over streaming data of high density. Conventional data and query indexing approaches generally do not apply for excessive costs in either maintenance or space. Aiming at these problems, this study first proposes a new indexing structure by fusing an adaptive cell and KDB-tree, namely CKDB-tree. A cell-tree indexing approach has been developed on the basis of the CKDB-tree that supports dynamic continuous queries. The approach significantly reduces the space costs and scales well with the increasing data size. Towards providing a scalable solution to filtering massive steaming data, this study has explored the feasibility to utilize the contemporary general-purpose computing on the graphics processing unit (GPGPU). The CKDB-tree-based approach has been extended to operate on both the CPU (host) and the GPU (device). The GPGPU-aided approach performs query indexing on the host while perform streaming data filtering on the device in a massively parallel manner. The two heterogeneous tasks execute in parallel and the latency of streaming data transfer between the host and the device is hidden. The experimental results indicate that (1) CKDB-tree can reduce the space cost comparing to the cell-based indexing structure by 60 percent on average, (2) the approach upon the CKDB-tree outperforms the traditional counterparts upon the KDB-tree by 66, 75 and 79 percent in average for uniform, skewed and hyper-skewed data in terms of update costs, and (3) the GPGPU-aided approach greatly improves the approach upon the CKDB-tree with the support of only a single Kepler GPU, and it provides real-time filtering of streaming data with 2.5M data tuples per second. The massively parallel computing technology exhibits great potentials in streaming data monitoring.

A scalable and fast OPTICS for clustering trajectory big data

Clustering trajectory data is an important way to mine hidden information behind moving object sampling data, such as understanding trends in movement patterns, gaining high popularity in geographic information and so on. In the era of `Big data', the current approaches for clustering trajectory data generally do not apply for excessive costs in both scalability and computing performance for trajectory big data. Aiming at these problems, this study first proposes a new clustering algorithm for trajectory big data, namely Tra-POPTICS by modifying a scalable clustering algorithm for point data (POPTICS). Tra-POPTICS has employed the spatiotemporal distance function and trajectory indexing to support trajectory data. Tra-POPTICS can process the trajectory big data in a distributed manner to meet a great scalability. Towards providing a fast solution to clustering trajectory big data, this study has explored the feasibility to utilize the contemporary general-purpose computing on the graphics processing unit (GPGPU). The GPGPU-aided clustering approach parallelized the Tra-POPTICS with the Hyper-Q feature of Kelper GPU and massive GPU threads. The experimental results indicate that (1) the Tra-POPTICS algorithm has a comparable clustering quality with T-OPTICS (the state of art work of clustering trajectories in a centralized fashion) and outperforms T-OPTICS by average four times in terms of scalability, and (2) the G-Tra-POPTICS has a comparable clustering quality with T-POPTICS as well and further gains about 30 speedup on average for clustering trajectories comparing to Tra-POPTICS with eight threads. The proposed algorithms exhibit great scalability and computing performance in clustering trajectory big data.

Towards adaptive synchronization measurement of large-scale non-stationary non-linear data

Synchronization measurement of non-stationary nonlinear data is an ongoing problem in the study of complex systems, e.g., neuroscience. Existing methods are largely based on Fourier transform and wavelet transform, and there is a lack of methods capable of (1) measuring the synchronization strength of multivariate data by adapting to non-stationary, non-linear dynamics, and (2) meeting the needs of sophisticated scientific or engineering applications. This study proposes an approach that measures the synchronization strength of bivariate non-stationary nonlinear data against phase differences. The approach (briefed as AD-PDSA) relies on adaptive algorithms for data decomposition. A parallelized approach was also developed with general-purpose computing on the graphics processing unit (GPGPU), which largely improved the scalability of data processing, namely, GAD-PDSA. We developed a model on the basis of GAD-PDSA to verify its effectiveness in analyzing multi-channel, event-related potential (ERP) recordings against Daubechies (DB) wavelet with reference to the Morlet wavelet transform (MWT). GAD-PDSA was applied to an EEG dataset obtained from epilepsy patients, and the synchronization analysis manifested an effective indicator of epileptic focus localization.

High throughput wavelet coherence analysis of neural series

The real–time estimation of coherence amongst neural signals from different brain areas is a critical issue in understanding brain functions. The wavelet coherence based on Monte Carlo method (MC–WTC) is effective in measuring the time–frequency coherence of neural signals, but it generates large intermediate data and could not be applied in real–time neural signal analysis. We develop a parallelised MC–WTC method with general–purpose computing on the graphics processing unit (GPGPU), namely G–MC–WTC, which speeds up the calculations using the CUDA toolkit. Simulation data showed that it can improve the runtime performance by almost 200 times. This method has been applied to a visual–auditory EEG data and to obtain the coherence information between different brain areas in real time. The result revealed a coherence difference in θ band at left temporal lobe. This method may become a useful tool for studying the cooperation mechanisms of brain regions in cognitive processes.

Global Synchronization Measurement of Multivariate Neural Signals with Massively Parallel Nonlinear Interdependence Analysis

The estimation of synchronization amongst multiple brain regions is a critical issue in understanding brain functions. There is a lack of an appropriate approach which is capable of 1) measuring the direction and strength of synchronization of activities of multiple brain regions, and 2) adapting to the quickly increasing sizes and scales of neural signals. Nonlinear Interdependence (NLI) analysis is an effective method for measuring synchronization direction and strength of bivariate neural signal. However, the method currently does not directly apply in handling multivariate signal. Its application in practice has also long been largely hampered by the ultra-high complexity of NLI algorithms. Aiming at these problems, this study 1) extends the conventional NLI to quantify the global synchronization of multivariate neural signals, and 2) develops a parallelized NLI method with general-purpose computing on the graphics processing unit (GPGPU), namely, G-NLI. The approach performs synchronization measurement in a massively parallel manner. The G-NLI has improved the runtime performance by more than 1000 times comparing to the original sequential NLI. Meanwhile, the G-NLI was employed to analyze 10-channel local field potential (LFP) recordings from a patient suffering from temporal lobe epilepsy. The results demonstrate that the proposed G-NLI method can support real-time global synchronization measurement and it could be successful in localization of epileptic focus.

Towards energy-efficient parallel analysis of neural signals

Recent advances of experimental methods and neuroscience research have made neural signals constantly massive and analysis of these signals highly compute-intensive. This study explore the possibility proposes a massively parallel approach for analysis of neural signals using General-purpose computing on the graphics processing unit (GPGPU). We demonstrate the uses and correctness of the proposed approach via a case of analyzing EEG with focal epilepsy. An experimental examination has been carried out to investigate (1) the GPGPU-aided approach's performance and (2) energy costs of the GPGPU-aided application versus the original CPU-only systems. Experimental results indicate that the proposed approach excels in both aspects.

GPGPU-Aided Ensemble Empirical-Mode Decomposition for EEG Analysis During Anesthesia

Ensemble empirical-mode decomposition (EEMD) is a novel adaptive time-frequency analysis method, which is particularly suitable for extracting useful information from noisy nonlinear or nonstationary data. Unfortunately, since the EEMD is highly compute-intensive, the method does not apply in real-time applications on top of commercial-off-the-shelf computers. Aiming at this problem, a parallelized EEMD method has been developed using general-purpose computing on the graphics processing unit (GPGPU), namely, G-EEMD. A spectral entropy facilitated by G-EEMD was, therefore, proposed to analyze the EEG data for estimating the depth of anesthesia (DoA) in a real-time manner. In terms of EEG data analysis, G-EEMD has dramatically improved the run-time performance by more than 140 times compared to the original serial EEMD implementation. G-EEMD also performs far better than another parallelized implementation of EEMD bases on conventional CPU-based distributed computing technology despite the latter utilizes 16 high-end computing nodes for the same computing task. Furthermore, the results obtained from a pharmacokinetics/pharmacodynamic (PK/PD) model analysis indicate that the EEMD method is slightly more effective than its precedent alternative method (EMD) in estimating DoA, the coefficient of determination R(2) by EEMD is significantly higher than that by EMD (p < 0.05, paired t-test) and the prediction probability P(k) by EEMD is also slighter higher than that by EMD (p < 0.2, paired t-test).

06/01417 Wavelet time-frequency analysis of accelerating and decelerating flows in a tube bank: Indrusiak, M. L. S. et al. Nuclear Engineering and Design, 2005, 235, (17–19), 1875–1887

Synchronization measurement of non-stationary nonlinear data is an ongoing problem in the study of complex systems, e.g., neuroscience. Existing methods are largely based on Fourier transform and wavelet transform, and there is a lack of methods capable of (1) measuring the synchronization strength of multivariate data by adapting to non-stationary, non-linear dynamics, and (2) meeting the needs of sophisticated scientific or engineering applications. This study proposes an approach that measures the synchronization strength of bivariate non-stationary nonlinear data against phase differences. The approach (briefed as AD-PDSA) relies on adaptive algorithms for data decomposition. A parallelized approach was also developed with general-purpose computing on the graphics processing unit (GPGPU), which largely improved the scalability of data processing, namely, GAD-PDSA. We developed a model on the basis of GAD-PDSA to verify its effectiveness in analyzing multi-channel, event-related potential (ERP) recordings against Daubechies (DB) wavelet with reference to the Morlet wavelet transform (MWT). GAD-PDSA was applied to an EEG dataset obtained from epilepsy patients, and the synchronization analysis manifested an effective indicator of epileptic focus localization.