Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “GPUs”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 307 records · Page 17

MARBLE: A Multi-GPU Aware Job Scheduler for Deep Learning on HPC Systems

Deep learning (DL) has become a key tool for solving complex scientific problems. However, managing the multi-dimensional large-scale data associated with DL, especially atop extant multiple graphics processing units (GPUs) in modern supercomputers poses significant challenges. Moreover, the latest high-performance computing (HPC) architectures bring different performance trends in training throughput compared to the existing studies. Existing DL optimizations such as larger batch size and GPU locality-aware scheduling have little effect on improving DL training throughput performance due to fast CPU-to-GPU connections. Additionally, DL training on multiple GPUs scales sublinearly. Thus, simply adding more GPUs to a system is ineffective. To this end, we design MARBLE, a first-of-its-kind job scheduler, which considers the non-linear scalability of GPUs at the intra-node level to schedule an appropriate number of GPUs per node for a job. By sharing the GPU resources on a node with multiple DL jobs, MARBLE avoids low GPU utilization in current multi-GPU DL training on HPC systems. Our comprehensive evaluation in the Summit supercomputer shows that MARBLE is able to improve DL training performance by up to 48.3% compared to the popular Platform Load Sharing Facility (LSF) scheduler. Compared to the state-of-the-art of DL scheduler, Optimus, MARBLE reduces the job completion time by up to 47%.

Han, Jingoo↗

FTTN: Feature-Targeted Testing for Numerical Properties of NVIDIA & AMD Matrix Accelerators

While NVIDIA has been the dominant provider of GPUs for HPC and ML, now AMD has several offerings of GPUs. This encourages programmers to try out AMD GPUs for new codes and also port existing codes over. Unfortunately, without understanding the floating-point differences between these GPU types, software development or porting can introduce bugs—and currently such an understanding is lacking. The magnitude of this open question becomes clear if one imagines the the number of floating-point precision choices (FP16, FP32, etc.), floating-point formats (standard floats, brain-float, etc.), and execution units available (elementary units, matrix/tensor cores, etc.) Questions such as rounding modes and subnormal support are also important. Most of these answers are unknown today or are hard to access. We provide the first testing-guided approach that answers a significant number of these questions. We also devise tests to reveal internal information (e.g., extra bits kept) to make sure that our findings are reliable. Many of our tests employ systematically generated random-programs, others apply fast-math flags and some involve fused multiplyadd. Especially for tensor/matrix cores, the tests have nontrivial logic that we present Our testing approach is reusable for the plethora of GPUs yet to be introduced. Our findings include up to 7 ulps of difference between NVIDIA and AMD for sin and cos at FP32 precision and 3 ulp at FP64. In our study of matrix cores (NVIDIA) and tensor cores (AMD), we have extensively characterized rounding modes (truncation versus round-to-nearest), the number of extra internal bits kept (whether 3 bits are kept or not), subnormal support for inputs and outputs across four different floating-point formats and across NVIDIA A100 and AMD MI250X GPUs. We believe that this wealth of data becoming available for the first time may help avoid significant porting bugs when migrating code across these platforms.

Li, Xinyi↗

GPU-Centric Communication on NVIDIA GPU Clusters with InfiniBand: A Case Study with OpenSHMEM

GPUs have become an essential component for building compute clusters with high compute density and high performance per watt. As such clusters scale to have 1000s of GPUs, efficiently moving data between the GPUs becomes imperative to get maximum performance. NVSHMEM is an implementation of the OpenSHMEM standard for NVIDIA GPU clusters which allows communication to be issued from inside GPU kernels. In earlier work, we have shown how NVSHMEM can be used to achieve better application performance on GPUs connected through PCIe or NVLink. As part of this effort, we implement IB verbs for Mellanox InfiniBand adapters in CUDA. We evaluate different design alternatives, taking into consideration the relaxed memory model, automatic memory access coalescing and thread hierarchy on the GPU. We also consider correctness issues that arise in these designs. We take advantage of these designs transparently or through API extensions in NVSHMEM. With micro-benchmarks, we show that a Nvidia Pascal P100 GPU is able saturate the network bandwidth using only one or two of its 56 available streaming multiprocessors (SM). On a single GPU using a single IB EDR adapter, we achieve a throughput of around 90 million messages per second. In addition, we implement a 2dstencil application kernel using NVSHMEM and compare its performance with a CUDA-aware MPI-based implementation that uses GPUDirect RDMA. Speedups in the range of 23% to 42% are seen for input sizes large enough to fill the occupancy of Nvidia Pascal P100 GPUs on 2 to 4 nodes indicating that there are gains to be had by eliminating the CPU from the communication path when all computation runs on the GPU.

Potluri, Sreeram↗

GMH: A Message Passing Toolkit for GPU Clusters

Driven by the market demand for high-definition 3D graphics, commodity graphics processing units (GPUs) have evolved into highly parallel, multi-threaded, many-core processors, which are ideal for data parallel computing. Many applications have been ported to run on a single GPU with tremendous speedups using general C-style programming languages such as CUDA. However, large applications require multiple GPUs and demand explicit message passing. This paper presents a message passing toolkit, called GMH (GPU Message Handler), on NVIDIA GPUs. This toolkit utilizes a data-parallel thread group as a way to map multiple GPUs on a single host to an MPI rank, and introduces a notion of virtual GPUs as a way to bind a thread to a GPU automatically. This toolkit provides high performance MPI style point-to-point and collective communication, but more importantly, facilitates event-driven APIs to allow an application to be managed and executed by the toolkit at runtime.

99 GENERAL AND MISCELLANEOUS↗

A graphics processing unit accelerated sparse direct solver and preconditioner with block low rank compression

We present the GPU implementation efforts and challenges of the sparse solver package STRUMPACK. The code is made publicly available on github with a permissive BSD license. STRUMPACK implements an approximate multifrontal solver, a sparse LU factorization which makes use of compression methods to accelerate time to solution and reduce memory usage. Multiple compression schemes based on rank-structured and hierarchical matrix approximations are supported, including hierarchically semi-separable, hierarchically off-diagonal butterfly, and block low rank. Here, in this paper, we present the GPU implementation of the block low rank (BLR) compression method within a multifrontal solver. Our GPU implementation relies on highly optimized vendor libraries such as cuBLAS and cuSOLVER for NVIDIA GPUs, rocBLAS and rocSOLVER for AMD GPUs and the Intel oneAPI Math Kernel Library (oneMKL) for Intel GPUs. Additionally, we rely on external open source libraries such as SLATE (Software for Linear Algebra Targeting Exascale), MAGMA (Matrix Algebra on GPU and Multi-core Architectures), and KBLAS (KAUST BLAS). SLATE is used as a GPU-capable ScaLAPACK replacement. From MAGMA we use variable sized batched dense linear algebra operations such as GEMM, TRSM and LU with partial pivoting. KBLAS provides efficient (batched) low rank matrix compression for NVIDIA GPUs using an adaptive randomized sampling scheme. The resulting sparse solver and preconditioner runs on NVIDIA, AMD and Intel GPUs. Interfaces are available from PETSc, Trilinos and MFEM, or the solver can be used directly in user code. We report results for a range of benchmark applications, using the Perlmutter system from NERSC, Frontier from ORNL, and Aurora from ALCF. For a high frequency wave equation on a regular mesh, using 32 Perlmutter compute nodes, the factorization phase of the exact GPU solver is about 6.5× faster compared to the CPU-only solver. The BLR-enabled GPU solver is about 13.8× faster than the CPU exact solver. For a collection of SuiteSparse matrices, the STRUMPACK exact factorization on a single GPU is on average 1.9× faster than NVIDIA’s cuDSS solver.

97 MATHEMATICS AND COMPUTING↗

Performance Results on CPU/GPU Exascale Architectures for OMEGA: The Ocean Model for E3SM Global Applications

The US Department of Energy (DOE) conducts climate simulations on some of the world’s largest supercomputers. These exascale machines use heterogeneous architectures with both CPUs and GPUs, and scientific codes must adapt to make full use of this computing power. Los Alamos National Lab is developing Omega: The Ocean Model for E3SM Global Applications, which is specifically designed for modern exascale computers. It uses external libraries that have been optimized for a variety of architectures to run on different supercomputers. Omega is an unstructured-mesh ocean model based on TRiSK numerical methods. It will be the new ocean component of the DOE’s Energy Exascale Earth System Model (E3SM). The algorithms in Omega follow those of the current ocean component, MPAS-Ocean, but it will be written in C++ rather than Fortran to take advantage of the Kokkos performance portability library. Omega spatial operators are written as Kokkos kernels to run efficiently on both CPUs and GPUs. Work on Omega began in 2023 with a new C++ framework for unstructured mesh partitioning, halo exchanges, parallel IO, and Kokkos interfaces. The current version, Omega-0, is being developed to solve the shallow water equations and at present includes all of the tendency terms but not time stepping. Here we share the results of Omega-0 verification and performance testing. Verification includes unit tests implemented with CTest as well as convergence tests in Polaris, an in-house python package with a large suite of test problems. Performance tests compare simulations conducted on CPUs versus GPUs and across different architectures: tests are run on Frontier, which has AMD “Optimized 3rd Gen EPYC” CPUs and AMD MI250X GPUs, as well as Perlmutter, which is composed of AMD EPYC 7763 CPUs and NVIDIA A100 GPUs.

58 GEOSCIENCES↗

Towards Microsecond Biological Molecular Dynamics Simulations on Hybrid Processors

Biomolecular simulations continue to become an increasingly important component of molecular biochemistry and biophysics investigations. Performance improvements in the simulations based on molecular dynamics (MD) codes are widely desired. This is particularly driven by the rapid growth of biological data due to improvements in experimental techniques. Unfortunately, the factors, which allowed past performance improvements of MD simulations, particularly the increase in microprocessor clock frequencies, are no longer improving. Hence, novel software and hardware solutions are being explored for accelerating the performance of popular MD codes. In this paper, we describe our efforts to port and optimize LAMMPS, a popular MD framework, on hybrid processors: graphical processing units (GPUs) accelerated multi-core processors. Our implementation is based on porting the computationally expensive, non-bonded interaction terms on the GPUs, and overlapping the computation on the CPU and GPUs. This functionality is built on top of message passing interface (MPI) that allows multi-level parallelism to be extracted even at the workstation level with the multi-core CPUs as well as extend the implementation on GPU clusters. The results from a number of typically sized biomolecular systems are provided and analysis is performed on 3 generations of GPUs from NVIDIA. Our implementation allows up to 30-40 ns/day throughput on a single workstation as well as significant speedup over Cray XT5, a high-end supercomputing platform. Moreover, detailed analysis of the implementation indicates that further code optimization and improvements in GPUs will allow {approx}100 ns/day throughput on workstations and inexpensive GPU clusters, putting the widely-desired microsecond simulation time-scale within reach to a large user community.

59 BASIC BIOLOGICAL SCIENCES↗

GPU-accelerated multitiered iterative phasing algorithm for fluctuation X-ray scattering

The multitiered iterative phasing (MTIP) algorithm is used to determine the biological structures of macromolecules from fluctuation scattering data. It is an iterative algorithm that reconstructs the electron density of the sample by matching the computed fluctuation X-ray scattering data to the external observations, and by simultaneously enforcing constraints in real and Fourier space. This paper presents the first ever MTIP algorithm acceleration efforts on contemporary graphics processing units (GPUs). The Compute Unified Device Architecture (CUDA) programming model is used to accelerate the MTIP algorithm on NVIDIA GPUs. The computational performance of the CUDA-based MTIP algorithm implementation outperforms the CPU-based version by an order of magnitude. Furthermore, the Heterogeneous-Compute Interface for Portability (HIP) runtime APIs are used to demonstrate portability by accelerating the MTIP algorithm across NVIDIA and AMD GPUs.

97 MATHEMATICS AND COMPUTING↗

Optimizing Data Movement for GPU-Based In-Situ Workflow Using GPUDirect RDMA

The extreme-scale computing landscape is increasingly dominated by GPU-accelerated systems. At the same time, in-situ workflows that employ memory-to-memory inter-application data exchanges have emerged as an effective approach for leveraging these extreme-scale systems. In the case of GPUs, GPUDirect RDMA enables third-party devices, such as network interface cards, to access GPU memory directly and has been adopted for intra-application communications across GPUs. In this paper, we present an interoperable framework for GPU-based in-situ workflows that optimizes data movement using GPUDirect RDMA. Specifically, we analyze the characteristics of the possible data movement pathways between GPUs from an in-situ workflow perspective, and design a strategy that maximizes throughput. Furthermore, we implement this approach as an extension of the DataSpaces data staging service, and experimentally evaluate its performance and scalability on a current leadership GPU cluster. The performance results show that the proposed design reduces data-movement time by up to 53% and 40% for the sender and receiver, respectively, and maintains excellent scalability for up to 256 GPUs.

Zhang, Bo↗

ExaCA: A performance portable exascale cellular automata application for alloy solidification modeling

Modeling the as-solidified grain structures that form during alloy processing is a critical component in understanding process-property relationships, particularly for additive manufacturing (AM) where grain structure is very sensitive to processing conditions. While cellular automata (CA)-based models have proven able to predict aspects of microstructure for several alloys and AM process conditions, long run times and large resource sets required limit the utility and the problem size to which existing CA models can be applied. As part of the ExaAM project, an initiative within the Exascale Computing Project (ECP) to develop, test, and optimize an exascale-capable coupled and self-consistent model of AM parts, we developed ExaCA (https://github.com/LLNL/ExaCA) for the liquid–solid phase transformation in the wake of AM melt pools. The CA-based code is parallelized using MPI and the Kokkos programming model, the latter enabling simulation on both CPUs and GPUs within a single-source implementation. Here, we detail the steps taken to transform a baseline, MPI-based CA code into one that is performant on CPUs and GPUs. Performance testing of ExaCA on Summit (a pre-exascale machine at Oak Ridge National Laboratory) was used to quantify CPU–GPU speedup comparing with equal numbers of nodes. Testing showed comparable CPU performance to the MPI-only CA code and a 5-20x speedup when running AM-based test problems using GPUs. The improved performance of CA through GPU utilization and the performance portable nature of ExaCA will enable accurate part-scale modeling by harnessing the power of current and future generations of high performance computing resources. Future work will include improving the strong scaling of ExaCA on GPUs by reducing load imbalance associated with the locality of the problem, and continuing performance optimization across exascale hardware.

36 MATERIALS SCIENCE↗

Development of a GPU-based high-performance radiative transfer model for the Infrared Atmospheric Sounding Interferometer (IASI)

Satellite-observed radiance is a nonlinear functional of surface properties and atmospheric temperature and absorbing gas profiles as described by the radiative transfer equation (RTE). In the era of hyperspectral sounders with thousands of high-resolution channels, the computation of the radiative transfer model becomes more time-consuming. The radiative transfer model performance in operational numerical weather prediction systems still limits the number of channels we can use in hyperspectral sounders to only a few hundreds. To take the full advantage of such high-resolution infrared observations, a computationally efficient radiative transfer model is needed to facilitate satellite data assimilation. In recent years the programmable commodity graphics processing unit (GPU) has evolved into a highly parallel, multi-threaded, many-core processor with tremendous computational speed and very high memory bandwidth. The radiative transfer model is very suitable for the GPU implementation to take advantage of the hardware's efficiency and parallelism where radiances of many channels can be calculated in parallel in GPUs. In this paper, we develop a GPU-based high-performance radiative transfer model for the Infrared Atmospheric Sounding Interferometer (IASI) launched in 2006 onboard the first European meteorological polar-orbiting satellites, METOP-A. Each IASI spectrum has 8461 spectral channels. The IASI radiative transfer model consists of three modules. The first module for computing the regression predictors takes less than 0.004% of CPU time, while the second module for transmittance computation and the third module for radiance computation take approximately 92.5% and 7.5%, respectively. Our GPU-based IASI radiative transfer model is developed to run on a low-cost personal supercomputer with four GPUs with total 960 compute cores, delivering near 4 TFlops theoretical peak performance. By massively parallelizing the second and third modules, we reached 364x speedup for 1 GPU and 1455x speedup for all 4 GPUs, both with respect to the original CPU-based single-threaded Fortran code with the -O{sub 2} compiling optimization. The significant 1455x speedup using a computer with four GPUs means that the proposed GPU-based high-performance forward model is able to compute one day's amount of 1,296,000 IASI spectra within nearly 10 min, whereas the original single CPU-based version will impractically take more than 10 days. This model runs over 80% of the theoretical memory bandwidth with asynchronous data transfer. A novel CPU-GPU pipeline implementation of the IASI radiative transfer model is proposed. The GPU-based high-performance IASI radiative transfer model is suitable for the assimilation of the IASI radiance observations into the operational numerical weather forecast model.

97 MATHEMATICS AND COMPUTING↗

Demonstrating GPU code portability and scalability for radiative heat transfer computations

High performance computing frameworks utilizing CPUs, Nvidia GPUs, and/or Intel Xeon Phis necessitate portable and scalable solutions for application developers. Nvidia GPUs in particular present numerous portability challenges with a different programming model, additional memory hierarchies, and partitioned execution units among streaming multiprocessors. This work presents modifications to the Uintah asynchronous many-task runtime and the Kokkos portability library to enable one single codebase for complex multiphysics applications to run across different architectures. Scalability and performance results are shown on multiple architectures for a globally coupled radiation heat transfer simulation, ranging from a single node to 16,384 Titan compute nodes.

Computer Science↗

Multi-GPU Implementation of a 3D Finite Difference Time Domain Earthquake Code on Heterogeneous Supercomputers

We have developed a highly scalable 3D Finite Difference GPU code for use in earthquake engineering and disaster management through regional petascale earthquake simulations. This MPI-CUDA code is based on a widely-used wave propagation code called AWP-ODC and restructured for high throughput and efficiency on a heterogeneous computing architecture. We present an effective communication reduction technique for leveraging GPUs with minimal PCI-e overhead, and a novel overlapping method to fully hide data communication latency between GPUs. The optimization concept used in this work can be extended to general stencil computing on a structured grid. The benchmarks demonstrated sustained 100 TFlops in single precision for 49 billion mesh points using 952 GPUs on the NCCS Titan Phase 5 system, which is a 77-fold speedup compared to the CPU version of the code. This multi-GPU implementation has been validated and used for a large-scale verification wave propagation simulation of Mw5.4 Chino Hills earthquake using 128 GPUs.

58 GEOSCIENCES↗

Scalable and accurate multi-GPU-based image reconstruction of large-scale ptychography data

Abstract While the advances in synchrotron light sources, together with the development of focusing optics and detectors, allow nanoscale ptychographic imaging of materials and biological specimens, the corresponding experiments can yield terabyte-scale volumes of data that can impose a heavy burden on the computing platform. Although graphics processing units (GPUs) provide high performance for such large-scale ptychography datasets, a single GPU is typically insufficient for analysis and reconstruction. Several works have considered leveraging multiple GPUs to accelerate the ptychographic reconstruction. However, most of these works utilize only the Message Passing Interface to handle the communications between GPUs. This approach poses inefficiency for a hardware configuration that has multiple GPUs in a single node, especially while reconstructing a single large projection, since it provides no optimizations to handle the heterogeneous GPU interconnections containing both low-speed (e.g., PCIe) and high-speed links (e.g., NVLink). In this paper, we provide an optimized intranode multi-GPU implementation that can efficiently solve large-scale ptychographic reconstruction problems. We focus on the maximum likelihood reconstruction problem using a conjugate gradient (CG) method for the solution and propose a novel hybrid parallelization model to address the performance bottlenecks in the CG solver. Accordingly, we have developed a tool, called PtyGer ( Pty chographic G PU(multipl e )-based r econstruction), implementing our hybrid parallelization model design. A comprehensive evaluation verifies that PtyGer can fully preserve the original algorithm’s accuracy while achieving outstanding intranode GPU scalability.

97 MATHEMATICS AND COMPUTING↗

CHOLLA: A NEW MASSIVELY PARALLEL HYDRODYNAMICS CODE FOR ASTROPHYSICAL SIMULATION

We present Computational Hydrodynamics On ParaLLel Architectures (Cholla ), a new three-dimensional hydrodynamics code that harnesses the power of graphics processing units (GPUs) to accelerate astrophysical simulations. Cholla models the Euler equations on a static mesh using state-of-the-art techniques, including the unsplit Corner Transport Upwind algorithm, a variety of exact and approximate Riemann solvers, and multiple spatial reconstruction techniques including the piecewise parabolic method (PPM). Using GPUs, Cholla evolves the fluid properties of thousands of cells simultaneously and can update over 10 million cells per GPU-second while using an exact Riemann solver and PPM reconstruction. Owing to the massively parallel architecture of GPUs and the design of the Cholla code, astrophysical simulations with physically interesting grid resolutions (≳256{sup 3}) can easily be computed on a single device. We use the Message Passing Interface library to extend calculations onto multiple devices and demonstrate nearly ideal scaling beyond 64 GPUs. A suite of test problems highlights the physical accuracy of our modeling and provides a useful comparison to other codes. We then use Cholla to simulate the interaction of a shock wave with a gas cloud in the interstellar medium, showing that the evolution of the cloud is highly dependent on its density structure. We reconcile the computed mixing time of a turbulent cloud with a realistic density distribution destroyed by a strong shock with the existing analytic theory for spherical cloud destruction by describing the system in terms of its median gas density.

79 ASTRONOMY AND ASTROPHYSICS↗

Portability for GPU-accelerated molecular docking applications for cloud and HPC: can portable compiler directives provide performance across all platforms?

High-throughput structure-based screening of drug-like molecules has become a common tool in biomedical research. Recently, acceleration with graphics processing units (GPUs) has provided a large performance boost for molecular docking programs. Both cloud and high-performance computing (HPC) resources have been used for large screens with molecular docking programs; while NVIDIA GPUs have dominated cloud and HPC resources, new vendors such as AMD and Intel are now entering the field, creating the problem of software portability across different GPUs. Ideally, software productivity could be maximized with portable programming models that are able to maintain high performance across architectures. While in many cases compiler directives have been used as an easy way to offload parallel regions of a CPU-based program to a GPU accelerator, they may also be an attractive programming model for providing portability across different GPU vendors, in which case the porting process may proceed in the reverse direction: from low-level, architecture-specific code to higher-level directive-based abstractions. MiniMDock is a new mini-application (miniapp) designed to capture the essential computational kernels found in molecular docking calculations, such as are used in phar-maceutical drug discovery efforts, in order to test different solutions for porting across GPU architectures. Here we extend MiniMDock to GPU offloading with OpenMP directives, and compare to performance of kernels using CUDA and HIP on NVIDIA and AMD GPUs, respectively, as well as across different compilers, exploring performance bottlenecks. We document this reverse-porting process, from highly optimized device code to a higher-level version using directives, compare code structure, and describe barriers that were overcome in this effort.

Thavappiragasam, Mathialakan↗

Understanding GPU errors on large-scale HPC systems and the implications for system design and operation

Increase in graphics hardware performance and improvements in programmability has enabled GPUs to evolve from a graphics-specific accelerator to a general-purpose computing device. Titan, the world's second fastest supercomputer for open science in 2014, consists of more dum 18,000 GPUs that scientists from various domains such as astrophysics, fusion, climate, and combustion use routinely to run large-scale simulations. Unfortunately, while the performance efficiency of GPUs is well understood, their resilience characteristics in a large-scale computing system have not been fully evaluated. We present a detailed study to provide a thorough understanding of GPU errors on a large-scale GPU-enabled system. Our data was collected from the Titan supercomputer at the Oak Ridge Leadership Computing Facility and a GPU cluster at the Los Alamos National Laboratory. We also present results from our extensive neutron-beam tests, conducted at Los Alamos Neutron Science Center (LANSCE) and at ISIS (Rutherford Appleron Laboratories, UK), to measure the resilience of different generations of GPUs. We present several findings from our field data and neutron-beam experiments, and discuss the implications of our results for future GPU architects, current and future HPC computing facilities, and researchers focusing on GPU resilience.

Tiwari, Devesh↗

Revisiting Huffman Coding: Toward Extreme Performance on Modern GPU Architectures

Today's high-performance computing (HPC) applications are producing vast volumes of data, which are challenging to store and transfer efficiently during the execution, such that data compression is becoming a critical technique to mitigate the storage burden and data movement cost. Huffman coding is arguably the most efficient Entropy coding algorithm in information theory, such that it could be found as a fundamental step in many modern compression algorithms such as DEFLATE. On the other hand, today's HPC applications are more and more relying on the accelerators such as GPU on supercomputers, while Huffman encoding suffers from low throughput on GPUs, resulting in a significant bottleneck in the entire data processing. In this paper, we propose and implement an efficient Huffman encoding approach based on modern GPU architectures, which addresses two key challenges: (1) how to parallelize the entire Huffman encoding algorithm, including codebook construction, and (2) how to fully utilize the high memory-bandwidth feature of modern GPU architectures. The detailed contribution is fourfold. (1) We develop an efficient parallel codebook construction on GPUs that scales effectively with the number of input symbols. (2) We propose a novel reduction based encoding scheme that can efficiently merge the codewords on GPUs. (3) We optimize the overall GPU performance by leveraging the state-of-the-art CUDA APIs such as Cooperative Groups. (4) We evaluate our Huffman encoder thoroughly using six real-world application datasets on two advanced GPUs and compare with our implemented multi-threaded Huffman encoder. Experiments show that our solution can improve the encoding throughput by up to 5.0x and 6.8x on NVIDIA RTX 5000 and V100, respectively, over the state-of-the-art GPU Huffman encoder, and by up to 3.3x over the multi-thread encoder on two 28-core Xeon Platinum 8280 CPUs.

Tian, Jiannan↗