SEARCH · Search NASA
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.
This project was a multi-partner CRADA with nine industrial partners. The CRADA with Xerox PARC made available the advanced quantum-level materials modeling (code name: PCGPP) code developed at LLNL to. the wider industrial community by porting and optimizing these codes for Cray Research supercomputers.
Abstract not provided.
Explore the source record for details and available documents.
Abstract not provided.
Explore the source record for details and available documents.
Abstract not provided.
Solid-state Drives (SSDs) have changed the landscape of storage systems and present a promising storage solution for data-intensive applications due to their low latency, high bandwidth, and low power consumption compared to traditional hard disk drives. SSDs achieve these desirable characteristics using internal parallelism—parallel access to multiple internal flash memory chips—and a Flash Translation Layer (FTL) that determines where data are stored on those chips so that they do not wear out prematurely. However, current state-of-the-art cache-based FTLs like the Demand-based Flash Translation Layer (DFTL) do not allow IO schedulers to take full advantage of internal parallelism, because they impose a tight coupling between the logical-to-physical address translation and the data access. In this study to address this limitation, we introduce a new FTL design called Parallel-DFTL that works with the DFTL to decouple address translation operations from data accesses. Parallel-DFTL separates address translation and data access operations into different queues, allowing the SSD to use concurrent flash accesses for both types of operations. We also present a Parallel-LRU cache replacement algorithm to improve the concurrency of address translation operations. To compare Parallel-DFTL against existing FTL approaches, we present a Parallel-DFTL performance model and compare its predictions against those for DFTL and an ideal page-mapping approach. We also implemented the Parallel-DFTL approach in an SSD simulator using real device parameters, and used trace-driven simulation to evaluate Parallel-DFTL’s efficacy. Our evaluation results show that Parallel-DFTL improved the overall performance by up to 32% for the real IO workloads we tested, and by up to two orders of magnitude with synthetic test workloads. Finally, we also found that Parallel-DFTL is able to achieve reasonable performance with a very small cache size and that it provides the best benefit for those workloads with large request size or with high write ratio.
Cloud platforms can provide flexible and cost-effective environments for parallel applications. However, the resource over-commitment issues, i.e., cloud providers often provide much more executable virtual CPUs than available physical CPUs, still impede the synchronization operations of parallel applications, causing severe performance degradation. Existing methods optimize parallel applications by promoting the priorities of involved VMs. They cannot fully explore the performance of parallel applications, because they ignore the time-slice requirements of different phases of parallel applications. Furthermore, non-parallel applications experience unsatisfied performance because of low scheduling priorities. Given empirical analysis on time-slices of virtual machines (VMs), we find that shortening time-slices can mitigate synchronization overhead which incurs during communication phases, while over-short time-slices cause frequent cache misses in computation phases. Accordingly, we propose an Adaptive Time-slice Control (ATC) mechanism. ATC first detects the phases of parallel applications based on lock latency or cache misses. Then, ATC shortens time-slices during communication phases and prolongs time-slices during computation phases for parallel applications, and sets a uniform time-slice for non-parallel applications. Finally, we evaluate ATC using seven well-known benchmarks with 25+ applications. Experiments show that ATC obtains 1.5-75x performance gain for running parallel applications than state-of-the-art solutions, with nearly unaffected impact on non-parallel applications.
When parallel algorithms for simulation were introduced in the 1970s, their development and use interested only experts in parallel computation. This circumstance changed as multi-core processors became commonplace, putting a parallel computer into the hands of every modeler. A natural outcome is growing interest in parallel simulation among persons not intimately familiar with parallel computing. At the same time, parallel simulation tools continue to be developed with the implicit assumption that the modeler is knowledgeable about parallel programming. The unintended consequence is a rapidly growing number of users of parallel simulation tools that are unlikely to recognize when the interaction of race conditions, partitioning strategies, and simultaneous action in their simulation models make results non-reproducible, thereby calling into question the validity of conclusions drawn from the simulation data. Here, we illustrate the potential dangers of exposing parallel algorithms to users who are not experts in parallel computation with example models constructed using existing parallel simulation tools. By doing so, we hope to refocus tool developers on usability, even if this new focus incurs loss of some performance.
Understanding parallel algorithms is crucial for accelerating scientific simulations on complex, distributed memory, high-performance computers. Modern algorithm classification approaches learn semantics directly from source code to differentiate between algorithms, however, accessing source code is not always possible. We can learn about parallel algorithms from observing their performance, as programs running the same algorithms and using the same hardware should exhibit similar performance characteristics. We present an approach to learn algorithm classes from parallel performance data directly in order to classify algorithms without access to the source code. We extend previous work to enable classifying parallel sorting algorithms using automatic instrumentation instead of requiring manual region annotations in the source code. In this work, we design and demonstrate a study for classification of parallel sorting algorithms using parallel performance data collected from automatic instrumentation, and evaluate the performance of our new methodology on classification. We leverage Caliper to collect the performance data, Thicket for our exploratory data analysis (EDA), and PyTorch and Scikit-learn to evaluate the effectiveness of random forests, support vector machines (SVMs), decision trees, neural networks, and logistic regressions on parallel performance data. Additionally, we study noise in parallel performance data, whether the removal of noise and pre-processing of the data is necessary to accurately classify parallel sorting algorithms, and determine the effectiveness of features created from performance data. In conclusion, we demonstrate classification accuracy for these five different models of up to 97.7% across four different parallel algorithm classes.
TorchBraid is a high-performance implementation of layer-parallel training for deep neural networks (DNNs) supporting MPI-based parallelism and GPU acceleration. Layer-parallel training has been developed to overcome the serialization inherent in forward and backward propagation of DNNs that limits utilization of computational resources in the strong scaling limit. To achieve this, TorchBraid integrates the PyTorch neural network framework with the state-of-the-art XBraid time-parallel library. Furthermore, this article presents the use and performance of TorchBraid, in addition to solutions for overcoming the algorithmic challenges inherent in combining automatic differentiation with layer-parallel. Results are presented with and without GPU acceleration for the Tiny ImageNet and MNIST image classification data sets, as well as recurrent neural networks. Overall, TorchBraid enables fast training of DNNs, both in a strong and weak scaling context. In addition to the TorchBraid software, several new advances in applying layer-parallel algorithms are detailed. Integration of layer-parallel with data-parallel algorithms is presented for the first time, showing the computational advantages of the combination. Standard deep learning techniques, like batch-normalization, are developed for layer-parallel training. Finally, a new approach combining layer-parallel with spatial coarsening in order to accelerate training for 3D image classification shows roughly a 10× speedup over serial execution.
GNU Parallel is a versatile and powerful tool for process parallelization widely used in scientific computing. This paper demonstrates its effective application in high-performance computing (HPC) environments, particularly focusing on its scalability and efficiency in executing large-scale high-throughput high-performance computing (HT-HPC) workflows. Through real-world examples, we highlight GNU Parallel’s performance across various HPC workloads, including GPU computing, container-based workloads, and node-local NVMe storage. Our results on two leading supercomputers, OLCF’s Frontier and NERSC’s Perlmutter, showcase GNU Parallel’s rapid process dispatching ability and its capacity to maintain low overhead even at extreme scales. We explore GNU Parallel’s application in massive parallel file transfers using a scheduled Data Transfer Node (DTN) cluster, emphasizing its broad utility in diverse scientific workflows. Beyond its direct application as a viable workflow manager, GNU Parallel can be employed in conjunction with other workflow systems as a "last-mile" parallelizing driver and as a quick prototyping tool to design and extract parallel profiles from application executions. We then argue that the potential for GNU Parallel to transform workflow management at extreme scales is substantial, paving the way for more efficient and effective scientific discoveries.
Modern hardware contains parallel execution resources that are well-suited for data-parallelism-vector units-and task parallelism-multicores. However, most work on parallel scheduling focuses on one type of hardware or the other. In this work, we present a scheduling framework that allows for a unified treatment of task- and data-parallelism. Our key insight is an abstraction, task blocks, that uniformly handles data-parallel iterations and task-parallel tasks, allowing them to be scheduled on vector units or executed independently as multicores. Our framework allows us to define schedulers that can dynamically select between executing task- blocks on vector units or multicores. We show that these schedulers are asymptotically optimal, and deliver the maximum amount of parallelism available in computation trees. To evaluate our schedulers, we develop program transformations that can convert mixed data- and task-parallel pro- grams into task block-based programs. Using a prototype instantiation of our scheduling framework, we show that, on an 8-core system, we can simultaneously exploit vector and multicore parallelism to achieve 14×-108× speedup over sequential baselines.
This work describes a space-time parallel algorithm with space-time adaptive mesh refinement (AMR). AMR with subcycling is added to multigrid reduction-in-time (MGRIT) in order to provide solution efficient adaptive grids with a reduction in work performed on coarser grids. This algorithm is achieved by integrating two software libraries: XBraid (Parallel time integration with multigrid. https://computation.llnl.gov/projects/parallel-timeintegration-multigrid) and Chombo (Chombo software package for AMR applications—design document, 2014). The former is a parallel time integration library using multigrid and the latter is a massively parallel structured AMR library. Employing this adaptive space-time parallel algorithm is Chord (Comput Fluids 123:202–217, 2015), a computational fluid dynamics (CFD) application code for solving compressible fluid dynamics problems. For the same solution accuracy, speedups are demonstrated from the use of space-time parallelization over the time-sequential integration on Couette flow and Stokes’ second problem. On a transient Couette flow case, at least a 1.5× speedup is achieved, and with a time periodic problem, a speedup of up to 13.7× over the time-sequential case is obtained. In both cases, the speedup is achieved by adding processors and exploring additional parallelization in time. The numerical experiments show the algorithm is promising for CFD applications that can take advantage of the time parallelism. Future work will focus on improving the parallel performance and providing more tests with complex fluid dynamics to demonstrate the full potential of the algorithm.
Energy efficiency of training and inferencing with large neural network models is a critical challenge facing the future of sustainable large-scale machine learning workloads. This paper introduces an alternative strategy, called phantom parallelism, to minimize the net energy consumption of traditional tensor (model) parallelism, the most energy-inefficient component of large neural network training. The approach is presented in the context of feed-forward network architectures as a preliminary, but comprehensive, proof-of-principle study of the proposed methodology. We derive new forward and backward propagation operators for phantom parallelism, implement them as custom autograd operations within an end-to-end phantom parallel training pipeline and compare its parallel performance and energy-efficiency against those of conventional tensor parallel training pipelines. Formal analyses that predict lower bandwidth and FLOP counts are presented with supporting empirical results on up to 256 GPUs that corroborate these gains. Experiments are shown to deliver ∼50% reduction in the energy consumed to train FFNs using the proposed phantom parallel approach when compared with conventional tensor parallel methods. Additionally, the proposed approach is shown to train smaller phantom models to the same model loss on smaller GPU counts as larger tensor parallel models on larger GPU counts offering the possibility for even greater energy savings.
We propose a new integrated method of exploiting model, batch and domain parallelism for the training of deep neural networks (DNNs) on large distributed-memory computers using minibatch stochastic gradient descent (SGD). Our goal is to find an efficient parallelization strategy for a fixed batch size using $P$ processes. Our method is inspired by the communication-avoiding algorithms in numerical linear algebra. We see $P$ processes as logically divided into a $P_r \times P_c$ grid where the $P_r$ dimension is implicitly responsible for model/domain parallelism and the $P_c$ dimension is implicitly responsible for batch parallelism. In practice, the integrated matrix-based parallel algorithm encapsulates these types of parallelism automatically. We analyze the communication complexity and analytically demonstrate that the lowest communication costs are often achieved neither with pure model nor with pure data parallelism. We also show how the domain parallel approach can help in extending the theoretical scaling limit of the typical batch parallel method.
Highlights: • Propose three parallel orbital-updating based plane-wave basis methods for electronic structure calculations. • These new methods can avoid the generating of large scale eigenvalue problems and then reduce the computational cost. • These new methods allow for two-level parallelization which is particularly interesting for large scale parallelization. • Numerical experiments show that these new methods are reliable and efficient for large scale calculations on modern supercomputers. - Abstract: Motivated by the recently proposed parallel orbital-updating approach in real space method , we propose a parallel orbital-updating based plane-wave basis method for electronic structure calculations, for solving the corresponding eigenvalue problems. In addition, we propose two new modified parallel orbital-updating methods. Compared to the traditional plane-wave methods, our methods allow for two-level parallelization, which is particularly interesting for large scale parallelization. Numerical experiments show that these new methods are more reliable and efficient for large scale calculations on modern supercomputers.
A second order kinetic model for turbulent ion parallel momentum transport is presented. A new nonresonant second order parallel momentum flux term is calculated. The resonant component of the ion parallel electrostatic force is the momentum source, while the nonresonant component of the ion parallel electrostatic force compensates for that of the nonresonant second order parallel momentum flux. The resonant component of the kinetic momentum flux can be divided into three parts, including the pinch term, the diffusive term, and the residual stress. By reassembling the pinch term and the residual stress, the residual stress can be considered as a pinch term of parallel wave-particle resonant velocity, and, therefore, may be called as “resonant velocity pinch” term. Considering the resonant component of the ion parallel electrostatic force is the transfer rate between resonant ions and waves (or, equivalently, nonresonant ions), a conservation equation of the parallel momentum of resonant ions and waves is obtained.