Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “gpu”

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 19 records

Work stealing for GPU-accelerated parallel programs in a global address space framework

Task parallelism is an attractive approach to automatically load balance the computation in a parallel system and adapt to dynamism exhibited by parallel systems. Exploiting task parallelism through work stealing has been extensively studied in shared and distributed-memory contexts. In this paper, we study the design of a system that uses work stealing for dynamic load balancing of task-parallel programs executed on hybrid distributed-memory CPU-graphics processing unit (GPU) systems in a global-address space framework. We take into account the unique nature of the accelerator model employed by GPUs, the significant performance difference between GPU and CPU execution as a function of problem size, and the distinct CPU and GPU memory domains. We consider various alternatives in designing a distributed work stealing algorithm for CPU-GPU systems, while taking into account the impact of task distribution and data movement overheads. These strategies are evaluated using microbenchmarks that capture various execution configurations as well as the state-of-the-art CCSD(T) application module from the computational chemistry domain

97 MATHEMATICS AND COMPUTING↗

Efficient Breadth First Search on Multi-GPU Systems Using GPU-Centric OpenSHMEM

NVSHMEM is an implementation of OpenSHMEM for NVIDIA GPUs which allows communication to be issued from inside CUDA kernels. In this work, we present an implementation of Breadth First Search for multi-GPU systems using NVSHMEM. We analyze the benefits and bottlenecks of moving fine-grained communication into CUDA kernels. Using our implementation of BFS, we achieve up to 75% improvement in performance compared to a CUDA-aware MPI-based implementation, in the best case.

97 MATHEMATICS AND COMPUTING↗

The Q continuum simulation: Harnessing the power of GPU accelerated supercomputers

Modeling large-scale sky survey observations is a key driver for the continuing development of high-resolution, large-volume, cosmological simulations. We report the first results from the “Q Continuum” cosmological N-body simulation run carried out on the GPU-accelerated supercomputer Titan. The simulation encompasses a volume of (1300 Mpc){sup 3} and evolves more than half a trillion particles, leading to a particle mass resolution of m{sub p}≃1.5⋅10{sup 8} M{sub ⊙}. At this mass resolution, the Q Continuum run is currently the largest cosmology simulation available. It enables the construction of detailed synthetic sky catalogs, encompassing different modeling methodologies, including semi-analytic modeling and sub-halo abundance matching in a large, cosmological volume. Here we describe the simulation and outputs in detail and present first results for a range of cosmological statistics, such as mass power spectra, halo mass functions, and halo mass-concentration relations for different epochs. We also provide details on challenges connected to running a simulation on almost 90% of Titan, one of the fastest supercomputers in the world, including our usage of Titan’s GPU accelerators.

79 ASTRONOMY AND ASTROPHYSICS↗

Parallel beamlet dose calculation via beamlet contexts in a distributed multi‐GPU framework

Purpose Dose calculation is one of the most computationally intensive, yet essential tasks in the treatment planning process. With the recent interest in automatic beam orientation and arc trajectory optimization techniques, there is a great need for more efficient model‐based dose calculation algorithms that can accommodate hundreds to thousands of beam candidates at once. Foundational work has shown the translation of dose calculation algorithms to graphical processing units ( GPU s), lending to remarkable gains in processing efficiency. But these methods provide parallelization of dose for only a single beamlet, serializing the calculation of multiple beamlets and under‐utilizing the potential of modern GPU s. In this paper, the authors propose a framework enabling parallel computation of many beamlet doses using a novel beamlet context transformation and further embed this approach in a scalable network of multi‐ GPU computational nodes. Methods The proposed context‐based transformation separates beamlet‐local density and TERMA into distinct beamlet contexts that independently provide sufficient data for beamlet dose calculation. Beamlet contexts are arranged in a composite context array with dosimetric isolation, and the context array is subjected to a GPU collapsed‐cone convolution superposition procedure, producing the set of beamlet‐specific dose distributions in a single pass. Dose from each context is converted to a sparse representation for efficient storage and retrieval during treatment plan optimization. The context radius is a new parameter permitting flexibility between the speed and fidelity of the dose calculation process. A distributed manager‐worker architecture is constructed around the context‐based GPU dose calculation approach supporting an arbitrary number of worker nodes and resident GPU s. Phantom experiments were executed to verify the accuracy of the context‐based approach compared to Monte Carlo and a reference CPU ‐ CCCS implementation for single beamlets and broad beams composed by addition of beamlets. Dose for representative 4 π beam sets was calculated in lung and prostate cases to compare its efficiency with that of an existing beamlet‐sequential GPU ‐ CCCS implementation. Code profiling was also performed to evaluate the scalability of the framework across many networked GPU s. Results The dosimetric accuracy of the context‐based method displays <1.35% and 2.35% average error from the existing serialized CPU ‐ CCCS algorithm and Monte Carlo simulation for beamlet‐specific PDD s in water and slab phantoms, respectively. The context‐based method demonstrates substantial speedup of up to two orders of magnitude over the beamlet‐sequential GPU ‐ CCCS method in the tested configurations. The context‐based framework demonstrates near linear scaling in the number of distributed compute nodes and GPU s employed, indicating that it is flexible enough to meet the performance requirements of most users by simply increasing the hardware utilization. Conclusions The context‐based approach demonstrates a new expectation of performance for beamlet‐based dose calculation methods. This approach has been successful in accelerating the dose calculation process for very large‐scale treatment planning problems ‐ such as automatic 4 π IMRT beam orientation and VMAT arc trajectory selection, with hundreds of thousands of beamlets ‐ in clinically feasible timeframes. The flexibility of this framework makes it as a strong candidate for use in a variety of other very large‐scale treatment planning tasks and clinical workflows.

Neph, Ryan↗

PPT-GPU: Scalable GPU Performance Modeling

Performance modeling is a challenging problem due to the complexities of hardware architectures. In this paper, we present PPT-GPU, a scalable and accurate simulation framework that enables GPU code developers and architects to predict the performance of applications in a fast, and accurate manner on different GPU architectures. PPT-GPU is part of the open source project, Performance Prediction Toolkit (PPT) developed at the Los Alamos National Laboratory. We extend the old GPU model in PPT that predict the runtimes of computational physics codes to offer better prediction accuracy, for which, we add models for different memory hierarchies found in GPUs and latencies for different instructions. To further show the utility of PPT-GPU, we compare our model against real GPU device(s) and the widely used cycle-accurate simulator, GPGPU-Sim using different workloads from RODINIA and Parboil benchmarks. The results indicate that the predicted performance of PPT-GPU is within a 10% error compared to the real device(s). In addition, PPT-GPU is highly scalable, where it is up to 450x faster than GPGPU-Sim with more accurate results.

97 MATHEMATICS AND COMPUTING↗

GPU accelerated Discrete Element Method (DEM) molecular dynamics for conservative, faceted particle simulations

Faceted shapes, such as polyhedra, are commonly found in systems of nanoscale, colloidal, and granular particles. Many interesting physical phenomena, like crystal nucleation and growth, vacancy motion, and glassy dynamics are challenging to model in these systems because they require detailed dynamical information at the individual particle level. Within the granular materials community the Discrete Element Method has been used extensively to model systems of anisotropic particles under gravity, with friction. We provide an implementation of this method intended for simulation of hard, faceted nanoparticles, with a conservative Weeks–Chandler–Andersen (WCA) interparticle potential, coupled to a thermodynamic ensemble. This method is a natural extension of classical molecular dynamics and enables rigorous thermodynamic calculations for faceted particles.

97 MATHEMATICS AND COMPUTING↗

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↗

Reliability Lessons Learned From GPU Experience With The Titan Supercomputer at Oak Ridge Leadership Computing Facility

The high computational capability of graphics processing units (GPUs) is enabling and driving the scientific discovery process at large-scale. The world’s second fastest supercomputer for open science, Titan, has more than 18,000 GPUs that computational scientists use to perform scientific simu- lations and data analysis. Understanding of GPU reliability characteristics, however, is still in its nascent stage since GPUs have only recently been deployed at large-scale. This paper presents a detailed study of GPU errors and their impact on system operations and applications, describing experiences with the 18,688 GPUs on the Titan supercom- puter as well as lessons learned in the process of efficient operation of GPUs at scale. These experiences are helpful to HPC sites which already have large-scale GPU clusters or plan to deploy GPUs in the future.

Tiwari, Devesh↗

Spy in the GPU-box: Covert and Side Channel Attacks on Multi-GPU System

The deep learning revolution has been enabled in large part by GPUs, and more recently accelerators, which make it possible to carry out computationally demanding training and inference in acceptable times. As the size of machine learning networks and workloads continues to increase, multi-GPU machines have emerged as an important platform offered on High Performance Computing and cloud data centers. Since these machines are shared among multiple users, it becomes increasingly important to protect applications against potential attacks. In this paper, we explore the vulnerability of Nvidia's DGX multi-GPU machines to covert and side channel attacks. These machines consist of a number of discrete GPUs that are interconnected through a combination of custom interconnect (NVLink) and PCIe connections. We reverse engineer the interconnected cache hierarchy and show that it is possible for an attacker on one GPU to cause contention on the L2 cache of another GPU. We use this observation to first develop a covert channel attack across two GPUs, achieving the best bandwidth of around 4 MB/s. We also develop a prime and probe attack on a remote GPU allowing an attacker to recover the cache access pattern of another workload. This access pattern can be used in any number of side channel attacks: we demonstrate a proof of concept attack that fingerprints the application running on the remote GPU, with high accuracy. We also develop a proof of concept attack to extract hyperparameters of a machine learning workload. Our work establishes for the first time the vulnerability of these machines to microarchitectural attacks and can guide future research to improve their security.

Dutta, Sankha↗

Panda: A Compiler Framework for Concurrent CPU $+$ GPU Execution of 3D Stencil Computations on GPU-accelerated Supercomputers

We present a new compiler framework for truly heterogeneous 3D stencil computation on GPU clusters. Our framework consists of a simple directive-based programming model and a tightly integrated source-to-source compiler. Annotated with a small number of directives, sequential stencil C codes can be automatically parallelized for large-scale GPU clusters. The most distinctive feature of the compiler is its capability to generate hybrid MPI$+$ CUDA$+$ OpenMP code that uses concurrent CPU$+$ GPU computing to unleash the full potential of powerful GPU clusters. The auto-generated hybrid codes hide the overhead of various data motion by overlapping them with computation. Test results on the Titan supercomputer and the Wilkes cluster show that auto-translated codes can achieve about 90 % of the performance of highly optimized handwritten codes, for both a simple stencil benchmark and a real-world application in cardiac modeling. The user-friendliness and performance of our domain-specific compiler framework allow harnessing the full power of GPU-accelerated supercomputing without painstaking coding effort.

97 MATHEMATICS AND COMPUTING↗

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.

Computer Science↗

Tartan: Evaluating Modern GPU Interconnect via a Multi-GPU Benchmark Suite

In this paper, we fill the gap by proposing a multi-GPU benchmark suite named Tartan, which contains microbenchmarks, scale-up and scale-out applications. We then apply Tartan to evaluate the four latest types of modern GPU interconnects, i.e., PCI- e, NVLink-V1, NVLink-V2 and InfiniBand with GPUDirect- RDMA from two recently released NVIDIA super AI platforms as well as ORNL’s exascale prototype system. Based on empirical evaluation, we observe four new types of NUMA effects: three types are triggered by NVLink’s topology, connectivity and routing, while one type is caused by PCI-e (i.e., anti-locality). They are very important for performance tuning in multi-GPU environment. Our evaluation results show that, unless the current CPU-GPU master-slave programming model can be replaced, it is difficult for scale-up multi-GPU applications to really benefit from faster intra-node interconnects such as NVLinks; while for inter-node scale-out applications, although interconnect is more crucial to the overall performance, GPUDirect-RDMA appears to be not always the optimal choice.

HPC↗

Fast plane wave density functional theory molecular dynamics calculations on multi-GPU machines

Plane wave pseudopotential (PWP) density functional theory (DFT) calculation is the most widely used method for material simulations, but its absolute speed stagnated due to the inability to use large scale CPU based computers. By a drastic redesign of the algorithm, and moving all the major computation parts into GPU, we have reached a speed of 12 s per molecular dynamics (MD) step for a 512 atom system using 256 GPU cards. This is about 20 times faster than the CPU version of the code regardless of the number of CPU cores used. Our tests and analysis on different GPU platforms and configurations shed lights on the optimal GPU deployments for PWP-DFT calculations. An 1800 step MD simulation is used to study the liquid phase properties of GaInP.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Novel features and GPU performance analysis for EM particle transport in the Celeritas code

Celeritas is a new computational transport code designed for high-performance simulation of high-energy physics detectors. This work describes some of its current capabilities and the design choices that enable the rapid development of efficient on-device physics. The abstractions that underpin the code design facilitate low-level performance tweaks that require no changes to the higher-level physics code. We evaluate a set of independent changes that together yield an almost 40% speedup over the original GPU code for a net performance increase of 220× for a single GPU over a single CPU running 8.4M tracks on a small demonstration physics app.

Johnson, Seth R.↗

Scaling Deep Learning on GPU and Knights Landing clusters

The speed of deep neural networks training has become a big bottleneck of deep learning research and development. For example, training GoogleNet by ImageNet dataset on one Nvidia K20 GPU needs 21 days. To speed up the training process, the current deep learning systems heavily rely on the hardware accelerators. However, these accelerators have limited on-chip memory compared with CPUs. To handle large datasets, they need to fetch data from either CPU memory or remote processors. We use both self-hosted Intel Knights Landing (KNL) clusters and multi-GPU clusters as our target platforms. From an algorithm aspect, current distributed machine learning systems are mainly designed for cloud systems. These methods are asynchronous because of the slow network and high fault-tolerance requirement on cloud systems. We focus on Elastic Averaging SGD (EASGD) to design algorithms for HPC clusters. Original EASGD used round-robin method for communication and updating. The communication is ordered by the machine rank ID, which is inefficient on HPC clusters. First, we redesign four efficient algorithms for HPC systems to improve EASGD's poor scaling on clusters. Async EASGD, Async MEASGD, and Hogwild EASGD are faster \textcolor{black}{than} their existing counterparts (Async SGD, Async MSGD, and Hogwild SGD, resp.) in all the comparisons. Finally, we design Sync EASGD, which ties for the best performance among all the methods while being deterministic. In addition to the algorithmic improvements, we use some system-algorithm codesign techniques to scale up the algorithms. By reducing the percentage of communication from 87% to 14%, our Sync EASGD achieves 5.3x speedup over original EASGD on the same platform. We get 91.5% weak scaling efficiency on 4253 KNL cores, which is higher than the state-of-the-art implementation.

60 APPLIED LIFE SCIENCES↗

RAP: Resource-aware Automated GPU Sharing for Multi-GPU Recommendation Model Training and Input Preprocessing

Ensuring high-quality recommendations for newly onboarded users requires the continuous retraining of Deep Learning Recommendation Models (DLRMs) with freshly generated data. To serve the online DLRM retraining, existing solutions use hundreds of CPU computing nodes designated for input preprocessing, causing significant power consumption that surpasses even the power usage of GPU trainers. To this end, we propose RAP, an end-to-end DLRM training framework that supports Resource-aware Automated GPU sharing for DLRM input Preprocessing and Training. The core idea of RAP is to accurately capture the remaining GPU computing resources during DLRM training for input preprocessing, achieving superior training efficiency without requiring additional resources. Specifically, RAP utilizes a co-running cost model to efficiently assess the costs of various input preprocessing operations, and it implements a resource-aware horizontal fusion technique that adaptively merges smaller kernels according to GPU availability, circumventing any interference with DLRM training. In addition, RAP leverages a heuristic searching algorithm that jointly optimizes both the input preprocessing graph mapping and the co-running schedule to maximize the end-to-end DLRM training throughput. The comprehensive evaluation shows that RAP achieves 78.3× speedup on average over CPU-based DLRM input preprocessing frameworks. In addition, the end-to-end training throughput of RAP is only 2.04% lower than the ideal case, which has no input preprocessing overhead.

Wang, Zheng↗