GPU-accelerated Version Research Articles

A simple graphics processing unit-accelerated propagation routine for laser pulses in the strong-field regime.

We present a simple and easy-to-implement Graphics Processing Unit (GPU)-accelerated routine to numerically simulate the propagation of ultrashort and intense laser pulses as they interact with a medium. The routine is based on the solution of Maxwell's wave equation in the frequency domain with an extended Crank-Nicolson algorithm implemented in the Nvidia CUDA C++ programming language. The main advantages of our method are its significant speed-up factor and its ease of implementation, requiring only basic knowledge of CUDA and C++. In this article, we review the strong-field wave equations to be solved and their discretization and demonstrate how to implement a numerical solver for them on an Nvidia GPU. We show the results of the simulation of a near-infrared laser pulse propagating through a partially ionized atomic gas and discuss the performance of our GPU-accelerated scheme. Compared to a naïve central processing unit implementation of the same routine, our GPU-accelerated version is up to 198 times faster in standard regimes.

CAREx: context-aware read extension of paired-end sequencing data

BackgroundCommonly used next generation sequencing machines typically produce large amounts of short reads of a few hundred base-pairs in length. However, many downstream applications would generally benefit from longer reads.ResultsWe present CAREx—an algorithm for the generation of pseudo-long reads from paired-end short-read Illumina data based on the concept of repeatedly computing multiple-sequence-alignments to extend a read until its partner is found. Our performance evaluation on both simulated data and real data shows that CAREx is able to connect significantly more read pairs (up to 99%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$99\\%$$\\end{document} for simulated data) and to produce more error-free pseudo-long reads than previous approaches. When used prior to assembly it can achieve superior de novo assembly results. Furthermore, the GPU-accelerated version of CAREx exhibits the fastest execution times among all tested tools.ConclusionCAREx is a new MSA-based algorithm and software for producing pseudo-long reads from paired-end short read data. It outperforms other state-of-the-art programs in terms of (i) percentage of connected read pairs, (ii) reduction of error rates of filled gaps, (iii) runtime, and (iv) downstream analysis using de novo assembly. CAREx is open-source software written in C++ (CPU version) and in CUDA/C++ (GPU version). It is licensed under GPLv3 and can be downloaded at (https://github.com/fkallen/CAREx).

Toward Horizon-scale Accretion onto Supermassive Black Holes in Elliptical Galaxies

We present high-resolution, three-dimensional hydrodynamic simulations of the fueling of supermassive black holes in elliptical galaxies from a turbulent medium on galactic scales, taking M87* as a typical case. The simulations use a new GPU-accelerated version of the Athena++ AMR code, and they span more than six orders of magnitude in radius, reaching scales similar to that of the black hole horizon. The key physical ingredients are radiative cooling and a phenomenological heating model. We find that the accretion flow takes the form of multiphase gas at radii less than about a kpc. The cold gas accretion includes two dynamically distinct stages: the typical disk stage in which the cold gas resides in a rotationally supported disk, and relatively rare chaotic stages (≲10% of the time) in which the cold gas inflows via chaotic streams. Though cold gas accretion dominates the time-averaged accretion rate at intermediate radii, accretion at the smallest radii is dominated by hot virialized gas at most times. The accretion rate scales with radius as when hot gas dominates, and we obtain near the event horizon, similar to what is inferred from EHT observations. The orientation of the cold gas disk can differ significantly on different spatial scales. We propose a subgrid model for accretion in lower-resolution simulations in which the hot gas accretion rate is suppressed relative to the Bondi rate by . Our results can also provide more realistic initial conditions for simulations of black hole accretion at the event horizon scale.

GPU Acceleration of Large-Scale Full-Frequency GW Calculations.

Many-body perturbation theory is a powerful method to simulate electronic excitations in molecules and materials starting from the output of density functional theory calculations. By implementing the theory efficiently so as to run at scale on the latest leadership high-performance computing systems it is possible to extend the scope of GW calculations. We present a GPU acceleration study of the full-frequency GW method as implemented in the WEST code. Excellent performance is achieved through the use of (i) optimized GPU libraries, e.g., cuFFT and cuBLAS, (ii) a hierarchical parallelization strategy that minimizes CPU-CPU, CPU-GPU, and GPU-GPU data transfer operations, (iii) nonblocking MPI communications that overlap with GPU computations, and (iv) mixed precision in selected portions of the code. A series of performance benchmarks has been carried out on leadership high-performance computing systems, showing a substantial speedup of the GPU-accelerated version of WEST with respect to its CPU version. Good strong and weak scaling is demonstrated using up to 25 920 GPUs. Finally, we showcase the capability of the GPU version of WEST for large-scale, full-frequency GW calculations of realistic systems, e.g., a nanostructure, an interface, and a defect, comprising up to 10 368 valence electrons.

Accelerated 2D Classification With ISAC Using GPUs.

A widely used approach to analyze single particles in electron microscopy data is 2D classification. This process is very computationally expensive, especially when large data sets are analyzed. In this paper we present GPU ISAC, a newly developed, GPU-accelerated version of the established Iterative Stable Alignment and Clustering (ISAC) algorithm for 2D images and generating class averages. While the previously existing implementation of ISAC relied on a computer cluster, GPU ISAC enables users to produce high quality 2D class averages from large-scale data sets on a single desktop machine equipped with affordable, consumer-grade GPUs such as Nvidia GeForce GTX 1080 TI cards. With only two such cards GPU ISAC matches the performance of twelve high end cluster nodes and, using high performance GPUs, is able to produce class averages from a million particles in between six to thirteen hours, depending on data set quality and box size. We also show GPU ISAC to scale linearly in all input dimensions, and thereby capable of scaling well with the increasing data load demand of future data sets. Further user experience improvements integrate GPU ISAC seamlessly into the existing SPHIRE GUI, as well as the TranSPHIRE on-the-fly processing pipeline. It is open source and can be downloaded at https://gitlab.gwdg.de/mpi-dortmund/sphire/cuISAC/

GPU-accelerated simulations of quantum annealing and the quantum approximate optimization algorithm

We study large-scale applications using a GPU-accelerated version of the massively parallel Jülich universal quantum computer simulator (JUQCS–G). First, we benchmark JUWELS Booster, a GPU cluster with 3744 NVIDIA A100 Tensor Core GPUs. Then, we use JUQCS–G to study the relation between quantum annealing (QA) and the quantum approximate optimization algorithm (QAOA). We find that a very coarsely discretized version of QA, termed approximate quantum annealing (AQA), performs surprisingly well in comparison to the QAOA. It can either be used to initialize the QAOA, or to avoid the costly optimization procedure altogether. Furthermore, we study the scaling of the success probability when using AQA for problems with 30 to 40 qubits. We find that the case with the largest discretization error scales most favorably, surpassing the best result obtained from the QAOA.

DeepGRP: engineering a software tool for predicting genomic repetitive elements using Recurrent Neural Networks with attention

BackgroundRepetitive elements contribute a large part of eukaryotic genomes. For example, about 40 to 50% of human, mouse and rat genomes are repetitive. So identifying and classifying repeats is an important step in genome annotation. This annotation step is traditionally performed using alignment based methods, either in a de novo approach or by aligning the genome sequence to a species specific set of repetitive sequences. Recently, Li (Bioinformatics 35:4408–4410, 2019) developed a novel software tool dna-brnn to annotate repetitive sequences using a recurrent neural network trained on sample annotations of repetitive elements.ResultsWe have developed the methods of dna-brnn further and engineered a new software tool DeepGRP. This combines the basic concepts of Li (Bioinformatics 35:4408–4410, 2019) with current techniques developed for neural machine translation, the attention mechanism, for the task of nucleotide-level annotation of repetitive elements. An evaluation on the human genome shows a 20% improvement of the Matthews correlation coefficient for the predictions delivered by DeepGRP, when compared to dna-brnn. DeepGRP predicts two additional classes of repeats (compared to dna-brnn) and is able to transfer repeat annotations, using RepeatMasker-based training data to a different species (mouse). Additionally, we could show that DeepGRP predicts repeats annotated in the Dfam database, but not annotated by RepeatMasker. DeepGRP is highly scalable due to its implementation in the TensorFlow framework. For example, the GPU-accelerated version of DeepGRP is approx. 1.8 times faster than dna-brnn, approx. 8.6 times faster than RepeatMasker and over 100 times faster than HMMER searching for models of the Dfam database.ConclusionsBy incorporating methods from neural machine translation, DeepGRP achieves a consistent improvement of the quality of the predictions compared to dna-brnn. Improved running times are obtained by employing TensorFlow as implementation framework and the use of GPUs. By incorporating two additional classes of repeats, DeepGRP provides more complete annotations, which were evaluated against three state-of-the-art tools for repeat annotation.

The Continental-Scale Soil-Moisture Precipitation Feedback in Europe with Parameterized and Explicit Convection

AbstractSoil moisture atmosphere interactions are key elements of the regional climate system. There is a well-founded hope that a more accurate representation of the soil moisture-precipitation feedback would improve the simulation of summer precipitation on daily to seasonal, to climate time scales. However, uncertainties have persistently remained as the simulated feedback is strongly sensitive to the model representation of deep convection. Here we assess the feedback representation using a GPU-accelerated version of the regional climate model COSMO. We simulate and compare the impact of continental-scale springtime soil-moisture anomalies on summer precipitation at convection-resolving (2.2 km) and convection-parameterizing resolution (12 km). We conduct re-analysis-driven simulations of 10 summer seasons (1999-2008) in continental Europe. While both simulations qualitatively agree on a positive sign of soil moisture-induced precipitation, they strongly differ in precipitation frequency: When convection is parameterized, wetter soil predominantly leads to more frequent precipitation events, and when convection is treated explicitly, they primarily become more intense. The results indicate that the sensitivity to soil moisture is stronger with parameterized convection, suggesting that the land surface-atmosphere coupling may be overestimated. In addition, the feedback’s sensitivity in complex terrain is assessed for soil perturbations of different horizontal scales. The convection-resolving simulations confirm a negative feedback for sub-continental soil moisture anomalies, which manifests itself in a local decrease of wet-hour frequency. However, the intensity feedback reinforces precipitation events at the same time (positive feedback). The two processes are represented differently in simulations with explicit and parameterized convection, explaining much of the difference between the two simulations.

Alfa Laval AB, Sweden

The tomographic reconstruction of biological specimens requires robust algorithms, able to deal with low density contrast and low element concentrations. At the IST/ITN microprobe facility new GPU-accelerated reconstruction software, JPIXET, has been developed, which can significantly increase the speed of quantitative reconstruction of Proton Induced X-ray Emission Tomography (PIXE-T) data. It has a user-friendly graphical user interface for pre-processing, data analysis and reconstruction of PIXE-T and Scanning Transmission Ion Microscopy Tomography (STIM-T). The reconstruction of PIXE-T data is performed using either an algorithm based on a GPU-accelerated version of the Maximum Likelihood Expectation Maximisation (MLEM) method or a GPU-accelerated version of the Discrete Image Space Reconstruction Algorithm (DISRA) (Sakellariou (2001) [2]). The original DISRA, its accelerated version, and the MLEM algorithm, were compared for the reconstruction of a biological sample of Caenorhabditis elegans – a small worm. This sample was analysed at the microbeam line of the AIFIRA facility of CENBG, Bordeaux. A qualitative PIXE-T reconstruction was obtained using the CENBG software package TomoRebuild (Habchi et al. (2013) [6]). The effects of pre-processing and experimental conditions on the elemental concentrations are discussed.

D3R Grand Challenge 4: prospective pose prediction of BACE1 ligands with AutoDock-GPU.

In this paper we describe our approaches to predict the binding mode of twenty BACE1 ligands as part of Grand Challenge 4 (GC4), organized by the Drug Design Data Resource. Calculations for all submissions (except for one, which used AutoDock4.2) were performed using AutoDock-GPU, the new GPU-accelerated version of AutoDock4 implemented in OpenCL, which features a gradient-based local search. The pose prediction challenge was organized in two stages. In Stage 1a, the protein conformations associated with each of the ligands were undisclosed, so we docked each ligand to a set of eleven receptor conformations, chosen to maximize the diversity of binding pocket topography. Protein conformations were made available in Stage 1b, making it a re-docking task. For all calculations, macrocyclic conformations were sampled on the fly during docking, taking the target structure into account. To leverage information from existing structures containing BACE1 bound to ligands available in the PDB, we tested biased docking and pose filter protocols to facilitate poses resembling those experimentally determined. Both pose filters and biased docking resulted in more accurate docked poses, enabling us to predict for both Stages 1a and 1b ligand poses within 2 Å RMSD from the crystallographic pose. Nevertheless, many of the ligands could be correctly docked without using existing structural information, demonstrating the usefulness of physics-based scoring functions, such as the one used in AutoDock4, for structure based drug design.

FullMonteCUDA: a fast, flexible, and accurate GPU-accelerated Monte Carlo simulator for light propagation in turbid media.

Optimizing light delivery for photodynamic therapy, quantifying tissue optical properties or reconstructing 3D distributions of sources in bioluminescence imaging and absorbers in diffuse optical imaging all involve solving an inverse problem. This can require thousands of forward light propagation simulations to determine the parameters to optimize treatment, image tissue or quantify tissue optical properties, which is time-consuming and computationally expensive. Addressing this problem requires a light propagation simulator that produces results quickly given modelling parameters. In previous work, we developed FullMonteSW: currently the fastest, tetrahedral-mesh, Monte Carlo light propagation simulator written in software. Additional software optimizations showed diminishing performance improvements, so we investigated hardware acceleration methods. This work focuses on FullMonteCUDA: a GPU-accelerated version of FullMonteSW which targets NVIDIA GPUs. FullMonteCUDA has been validated across several benchmark models and, through various GPU-specific optimizations, achieves a 288-936x speedup over the single-threaded, non-vectorized version of FullMonteSW and a 4-13x speedup over the highly optimized, hand-vectorized and multi-threaded version. The increase in performance allows inverse problems to be solved more efficiently and effectively.

A GPU accelerated discontinuous Galerkin incompressible flow solver

We present a GPU-accelerated version of a high-order discontinuous Galerkin discretization of the unsteady incompressible Navier–Stokes equations. The equations are discretized in time using a semi-implicit scheme with explicit treatment of the nonlinear term and implicit treatment of the split Stokes operators. The pressure system is solved with a conjugate gradient method together with a fully GPU-accelerated multigrid preconditioner which is designed to minimize memory requirements and to increase overall performance. A semi-Lagrangian subcycling advection algorithm is used to shift the computational load per timestep away from the pressure Poisson solve by allowing larger timestep sizes in exchange for an increased number of advection steps. Numerical results confirm we achieve the design order accuracy in time and space. We optimize the performance of the most time-consuming kernels by tuning the fine-grain parallelism, memory utilization, and maximizing bandwidth. To assess overall performance we present an empirically calibrated roofline performance model for a target GPU to explain the achieved efficiency. We demonstrate that, in the most cases, the kernels used in the solver are close to their empirically predicted roofline performance.

GPU Accelerated Computation of Isotropic Chemical Shifts Offers New Dimension of Structure Refinement in Largescale Molecular Dynamics Simulation

Semi-empirical chemical shift prediction, where molecular coordinates and empirical data are reduced to backbone and sidechain chemical shifts through input to differentiable functions, offers a unique opportunity to validate and refine protein structure during largescale molecular dynamics (MD) simulation but poses the challenge of optimizing performance to levels comparable to modern GPU-enabled MD engines. The software PPM_One predicts backbone and sidechain chemical shifts with competitive accuracy to purely data-driven counterparts while obviating the need for deep network processing and homology assignment. To poise PPM_One for practical application in largescale biomolecular modeling, a newly optimized GPU-accelerated version of PPM_One was developed. Code refactoring and parallel GPU-processing through the directive-based OpenACC API provide a 67-fold speedup between unaccelerated and Nvidia V100 accelerated PPM_One test-cases, comparatively, in a system of 21 million protein atoms. The core functions that comprise the prediction model and compute perturbation due to Hydrogen bonding, magnetic anisotropy and ring currents were profiled at 271x, 322x, and 211x speedups, respectively. Remaining components of the predictive model saw similar individual speedups through profiling. Implementing the newly optimized functional core into NAMD is NMRForces, a global force module that enables users to compute backbone and sidechain chemical shifts at each timestep and force atoms along the prediction model's gradient with a magnitude proportional to deviance from experimental chemical shift data.

New Binding Mode of SLURP Protein to a7 Nicotinic Acetylcholine Receptor Revealed by Computer Simulations

SLURP-1 is a member of three-finger toxin-like proteins. Their characteristic feature is a set of three beta strands extruding from hydrophobic core stabilized by disulfide bonds. Each beta-strand carries a flexible loop, which is responsible for recognition. SLURP-1 was recently shown to act as an endogenous growth regulator of keratinocytes and tumor suppressor by reducing cell migration and invasion by antagonizing the pro-malignant effects of nicotine. This effect is achieved through allosteric interaction with alpha7 nicotinic acetylcholine receptors (alpha-7 nAChRs) in an antagonist-like manner. Moreover, this interaction is unaffected by several well-known agents specifically alpha-bungarotoxin. In this work, we carry out the conformational analysis of the SLURP-1 by a microsecond-long full-atom explicit solvent molecular dynamics simulations followed by clustering, to identify representative states. To achieve this timescale we employed a GPU-accelerated version of GROMACS modeling package. To avoid human bias in clustering we used a non-parametric clustering algorithm Affinity Propagation adapted for biomolecules and HPC environments. Then, we applied protein-protein molecular docking of the ten most massive clusters to alpha7-nAChRs in order to test if structural variability can affect binding. Docking simulations revealed the unusual binding mode of one of the minor SLURP-1 conformations.

GPU acceleration and performance of the particle-beam-dynamics code Elegant

Elegant is an accelerator physics and particle-beam dynamics code widely used for modeling and design of a variety of high-energy particle accelerators and accelerator-based systems. In this paper we discuss a recently developed version of the code that can take advantage of CUDA-enabled graphics processing units (GPUs) to achieve significantly improved performance for a large class of simulations that are important in practice. The GPU version is largely defined by a framework that simplifies implementations of the fundamental kernel types that are used by Elegant: particle operations, reductions, particle loss, histograms, array convolutions and random number generation. Accelerated performance on the Titan Cray XK-7 supercomputer is approximately 6–10 times better with the GPU than all the CPU cores associated with the same node count. In addition to performance, the maintainability of the GPU-accelerated version of the code was considered a key design objective. Accuracy with respect to the CPU implementation is also a core consideration. Four different methods are used to ensure that the accelerated code faithfully reproduces the CPU results. Program summaryProgram Title: Kernels from the GPU-accelerated ElegantProgram Files doi:http://dx.doi.org/10.17632/jc465zy7p5.1Licensing provisions: MITProgramming language: C/C++/CUDANature of problem: The original design of the Elegant accelerator physics code was implemented on central processing units with message-passing interface parallelization. This implementation is not able to use next-generation multicore systems.Solution method: In this package we develop routines based on the CUDA language extensions to C++ that enable porting the Elegant code to be run on graphics processing units (GPUs). Special consideration is given to algorithms that require collective communication on the GPU.Additional comments including restrictions and unusual features: The full Elegant source code is freely available from Argonne National Laboratory and these distributions include the GPU code in the later releases.

Block Conjugate-Gradient Method With Multilevel Preconditioning and GPU Acceleration for FEM Problems in Electromagnetics

In this letter, a graphics processing unit (GPU)-accelerated block conjugate-gradient solver with multilevel preconditioning is presented for solving a large system of sparse equations with multiple right-hand sides (RHSs), which arise in the finite-element analysis of electromagnetic problems. We demonstrate that blocking reduces the time to solution significantly and allows for better utilization of the computing power of GPUs, especially when the system matrix is complex valued. We show that, for a complex-valued sparse matrix with over $\text{1.2}$ million unknowns and 16 RHSs, the block version of the preconditioned conjugate-gradient method using a single NVIDIA Tesla P100 accelerator is almost three times faster than a nonblocked version. Numerical tests have also shown that a central processing unit (CPU)-only nonblocked complex implementation is unlikely to benefit much from blocking. Compared to an optimized CPU-only solver using an Intel Xeon E5-2680 v3 with 12 cores, the nonblocked GPU-accelerated version was $\text{4.5}$ times faster, while the blocked version was $\text{12.4}$ times faster.

Accelerating Computer-Based Recognition of Fynbos Leaves Using a Graphics Processing Unit

The Cape Floristic Kingdom (CFK) is the most diverse floristic kingdom in the world and has been declared an international heritage site. However, it is under threat from wild fires and invasive species. Much of the work of managing this natural resource, such as removing alien vegetation or fighting wild fires, is done by volunteers and casual workers. Many fynbos species, for which the Table Mountain National Park is known, are difficult to identify, particularly by non-expert volunteers. Accurate and fast identification of plant species would be beneficial in these contexts. The Fynbos Leaf Optical Recognition Application (FLORA) was thus developed to assist in the recognition of plants of the CFK. The first version of FLORA was developed as a rapid prototype in MATLAB; it utilized sequential algorithms to identify plant leaves, and much of this code was interpreted M files. The initial implementation suffered from slow performance, though, and could not run as a lightweight standalone executable, making it cumbersome. FLORA was thus re-developed as a standalone C++ version that was subsequently enhanced further by accelerating critical routines, by running them on a graphics processing unit (GPU). This paper presents the design and testing of both the C++ version and the GPU-accelerated version of FLORA. Comparative testing was done on all three versions of FLORA, viz., the original MATLAB prototype, the C++ non-accelerated version, and the C++ GPU-accelerated version to show the performance and accuracy of the different versions. The accuracy of the predictions remained consistent across versions. The C++ version was noticeable faster than the original prototype, achieving an average speed-up of 8.7 for high-resolution 3456x2304 pixel images. The GPU-accelerated version was even faster, saving 51.85 ms on average for high-resolution images. Such a time saving would be perceptible for batch processing, such as rebuilding feature descriptors for all the leaves in the leaf database. Further work on this project involves testing the system with a wider variety of leaves and trying different machine learning algorithms for the leaf prediction routines.

A GPU-Accelerated Fourth-Order Runge–Kutta in the Interaction Picture Method for the Simulation of Nonlinear Signal Propagation in Multimode Fibers

The nonlinear signal propagation in fibers can be described by the nonlinear Schrodinger equation and the Manakov equation. Most commonly, split-step Fourier methods (SSFM) are applied to solve these nonlinear equations. The numerical simulation of the nonlinear signal propagation is especially challenging for multimode fibers, particularly if the calculation of very small step sizes or a large number of steps is required. Instead of utilizing SSFM, the fourth-order Runge–Kutta in the Interaction Picture (RK4IP) method can be applied. This method has the potential to reduce the numerical error while simultaneously allowing an increased step size. These advantages come at the price of a higher numerical effort compared to the SSFM method for the same step size. Since the simulation of the signal propagation in multimode fibers is already quite challenging, parallelization becomes an even more interesting option. We demonstrate the adaptation of the RK4IP method to simulate the nonlinear signal propagation in multimode fibers, including its parallelization. Besides comparing the performance of a parallelized implementation for multicore CPUs and a GPU-accelerated version, we discuss efficient strategies to implement the RK4IP method on a GPU accelerator with CUDA. In addition, the RK4IP implementation is numerically compared with a conventional SSFM implementation.

GPU-accelerated Tersoff potentials for massively parallel Molecular Dynamics simulations

The Tersoff potential is one of the empirical many-body potentials that has been widely used in simulation studies at atomic scales. Unlike pair-wise potentials, the Tersoff potential involves three-body terms, which require much more arithmetic operations and data dependency. In this contribution, we have implemented the GPU-accelerated version of several variants of the Tersoff potential for LAMMPS, an open-source massively parallel Molecular Dynamics code. Compared to the existing MPI implementation in LAMMPS, the GPU implementation exhibits a better scalability and offers a speedup of 2.2X when run on 1000 compute nodes on the Titan supercomputer. On a single node, the speedup ranges from 2.0 to 8.0 times, depending on the number of atoms per GPU and hardware configurations. The most notable features of our GPU-accelerated version include its design for MPI/accelerator heterogeneous parallelism, its compatibility with other functionalities in LAMMPS, its ability to give deterministic results and to support both NVIDIA CUDA- and OpenCL-enabled accelerators. Our implementation is now part of the GPU package in LAMMPS and accessible for public use.

Multilayer Packet Classification With Graphics Processing Units

The rapid growth of server virtualization has ignited a wide adoption of software-based virtual switches, with significant interest in speeding up their performance. In a similar trend, software-defined networking (SDN), with its strong reliance on rule-based flow classification, has also created renewed interest in multi-dimensional packet classification. However, despite these recent advances, the performance of current software-based packet classifiers is still limited, mostly by the low parallelism of general-purpose CPUs. In this paper, we explore how to accelerate packet classification using the high parallelism and latency-hiding capabilities of graphic processing units (GPUs). We implement GPU-accelerated versions for both linear and tuple search, currently deployed in virtual switches, and also introduce a novel algorithm called Bloom search. These algorithms are integrated with high-speed packet I/O to build GSwitch, a GPU-accelerated software switch, and also to extend Open vSwitch. Our experimental evaluation indicates that, under realistic rule sets, GSwitch is at least 7 ${\times}$ faster than an equally-priced CPU classifier. We also show that our GPU-accelerated Open vSwitch outperforms the classic Open vSwitch implementation by a factor of 10, on average.