SEARCH · Engineering Papers
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.
SAIL, a high level ALGOL language for the PDP-10, is extended to operate under the TENEX time sharing system without executing DEC system calls. A large set of TENEX oriented runtime routines are added to allow complete access to TENEX. The emphasis is on compatibility of programs across time sharing systems and integrity of the language.
Here, we present a sparse linear system solver that is based on a multifrontal variant of Gaussian elimination and exploits low-rank approximation of the resulting dense frontal matrices. We use hierarchically semiseparable (HSS) matrices, which have low-rank off-diagonal blocks, to approximate the frontal matrices. For HSS matrix construction, a randomized sampling algorithm is used together with interpolative decompositions. The combination of the randomized compression with a fast ULV HSS factoriz ation leads to a solver with lower computational complexity than the standard multifrontal method for many applications, resulting in speedups up to 7 fold for problems in our test suite. The implementation targets many-core systems by using task parallelism with dynamic runtime scheduling. Numerical experiments show performance improvements over state-of-the-art sparse direct solvers. The implementation achieves high performance and good scalability on a range of modern shared memory parallel systems, including the Intel Xeon Phi (MIC). The code is part of a software package called STRUMPACK - STRUctured Matrices PACKage, which also has a distributed memory component for dense rank-structured matrices.
Programming models intended to run on exascale systems have a number of challenges to overcome, specially the sheer size of the system as measured by the number of concurrent software entities created and managed by the underlying runtime. It is clear from the size of these systems that any state maintained by the programming model has to be strictly sub-linear in size, in order not to overwhelm memory usage with pure overhead. A principal feature of Partitioned Global Address Space (PGAS) models is providing easy access to global-view distributed data structures. In order to provide efficient access to these distributed data structures, PGAS models must keep track of metadata such as where array sections are located with respect to processes/threads running on the HPC system. As PGAS models and applications become ubiquitous on very large transpetascale systems, a key component to their performance and scalability will be efficient and judicious use of memory for model overhead (metadata) compared to application data. We present an evaluation of several strategies to manage PGAS metadata that exhibit different space/time tradeoffs. We use two real-world PGAS applications to capture metadata usage patterns and gain insight into their communication behavior.
Scalable deep neural network training has been gaining prominence because of the increasing importance of deep learning in a multitude of scientific and commercial domains. Consequently, a number of researchers have investigated techniques to optimize deep learning systems. Much of the prior work has focused on runtime and algorithmic enhancements to optimize the computation and communication. Despite these enhancements, however, deep learning systems still suffer from scalability limitations, particularly with respect to data I/O. This situation is especially true for training models where the computation can be effectively parallelized, leaving I/O as the major bottleneck. In fact, our analysis shows that I/O can take up to 90% of the total training time. Thus, in this article, we first analyze LMDB, the most widely used I/O subsystem of deep learning frameworks, to understand the causes of this I/O inefficiency. Based on our analysis, we propose LMDBIO—an optimized I/O plugin for scalable deep learning. LMDBIO includes six novel optimizations that together address the various shortcomings in existing I/O for deep learning. Lastly, our experimental results show that LMDBIO significantly outperforms LMDB in all cases and improves overall application performance by up to 65-fold on a 9,216-core system.
Scientific workflows consist of multiple, connected applications, with data and results flowing from one to another in a pipeline. Traditionally, such workflows are executed in sequential order, storing intermediate data in storage disks. Co-scheduling application workflows concurrently on the same compute nodes would greatly reduce the cost of moving data to/from storage and allow real-time analysis of intermediate results. Nevertheless, most parallel programming runtimes do not allow seamless integration of various applications in a scientific workflow, in part due to the complexity of managing data and resources. The situation is even more complicated for heterogeneous systems. In this work we extend the Minos Computing Library (MCL) runtime to accelerate pipe-lined and parallel workloads where multiple applications are running in the same system. MCL’s asynchronous task library and runtime dynamically manages resources to allow co-scheduling of multiple processes sharing heterogeneous resources. In addition, we design a custom ex- tension of the Open Compute Language (OpenCL) to enable multiple processes to share device memory. We enable MCL to coordinate these shared buffers to allow for easy, fast data sharing between applications. Using malleable micro-benchmarks and two application workflows that combine scientific simulation and AI-based analysis, we show that our method outperforms traditional approaches.
The continual expansion of high-performance computing (HPC) brings with it an increasing need for efficiency. Heavy investment in energy, hardware, and software infrastructure to support peta- and exascale computing requires the optimization of existing systems and, wherever possible, the discernment and adoption of best-practices towards these goals. Such is the case for runtime prediction. When a job is submitted to an HPC system, an estimate of its runtime is provided by the user in the form of "requested wallclock". Error in this user-provided estimate can lead to jobs being prematurely killed by the scheduler, increased wait time on the queue, and decreased system utilization. More than fifteen years of research has been directed at mitigating these effects by using data-driven runtime predictions. Codified here is a set of commonalities and insights emerging from this body of work, which we present as recommendations and best practices. These practices are combined into a methodological approach described and evaluated on an 11-million-job dataset from the National Renewable Energy Laboratory's petascale HPC system, Eagle. This dataset and the accompanying codebase have been released to the public domain for the benefit of the wider HPC research community.
The Space Station Data Management System (DMS), associated computing subsystems, and applications have varying degrees of reliability associated with their operation. A model has been developed (McKay '86) which allows the DMS runtime environment to appear as an Ada virtual machine to applications executing within it. This model is modular, flexible, and dynamically configurable to allow for evolution and growth over time. Support for Fault-tolerant computing is included within this model. The basic primitive involved in this support is based on atomic actions (Grey '78). An atomic action possesses two fundamental properties: (1) it is indivisible with respect to concurrent actions, and (2) it is indivisible with respect to failure. A transaction is a collection of atomic actions which collectively appear to be one action. Transactions may be nested, providing even more powerful support for reliability. A proposed approach is described for providing support for nested atomic transactions within the Ada runtime model developed for the Space Station environment. The level of support is modular, flexible and dynamically configurable just like the overall runtime support environment.
The continual expansion of high-performance computing (HPC) brings with it an increasing need for efficiency. Heavy investment in energy, hardware, and software infrastructure to support peta- and exascale computing requires the optimization of existing systems and, wherever possible, the discernment and adoption of best-practices towards these goals. Such is the case for runtime prediction. When a job is submitted to an HPC system, an estimate of its runtime is provided by the user in the form of "requested wallclock''. Error in this user-provided estimate can lead to jobs being prematurely killed by the scheduler, increased wait time on the queue, and decreased system utilization. More than fifteen years of research has been directed at mitigating these effects by using data-driven runtime predictions. Codified here is a set of commonalities and insights emerging from this body of work, which we present as recommendations and best practices. These practices are combined into a methodological approach described and evaluated on an 11-million-job dataset from the National Renewable Energy Laboratory's petascale HPC system, Eagle. This dataset and the accompanying codebase have been released to the public domain for the benefit of the wider HPC research community.
Supercomputers complexity continue to rapidly increase, requiring novel software solutions to aid in addressing the extreme abundance of parallelism, restrictive power constraints, and fault tolerance requirements. While industrial standards like MPI and OpenMP have undergone extensive efforts to prepare for future systems [1, 2], ne grain asynchronous runtimes have demonstrated an innate agility in extracting performance which is otherwise arduous to attain with current execution models
Cloning is a technique to efficiently simulate a tree of multiple what-if scenarios that are unraveled during the course of a base simulation. However, cloned execution is highly challenging to realize on large, distributed memory computing platforms, due to the dynamic nature of the computational load across clones, and due to the complex dependencies spanning the clone tree. In this paper, we present the conceptual simulation framework, algorithmic foundations, and runtime interface of CloneX, a new system we designed for scalable simulation cloning. It efficiently and dynamically creates whole logical copies of a dynamic tree of simulations across a large parallel system without full physical duplication of computation and memory. The performance of a prototype implementation executed on up to 1,024 graphical processing units of a supercomputing system has been evaluated with three benchmarks—heat diffusion, forest fire, and disease propagation models—delivering a speed up of over two orders of magnitude compared to replicated runs. Finally, the results demonstrate a significantly faster and scalable way to execute many what-if scenario ensembles of large simulations via cloning using the CloneX interface.
Explore the source record for details and available documents.
Emissions are a central component of atmospheric chemistry models. The Harmonized Emissions Component (HEMCO) is a software component for computing emissions from a user-selected ensemble of emission inventories and algorithms. It allows users to re-grid, combine, overwrite, subset, and scale emissions from different inventories through a configuration file and with no change to the model source code. The configuration file also maps emissions to model species with appropriate units. HEMCO can operate in offline stand-alone mode, but more importantly it provides an online facility for models to compute emissions at runtime. HEMCO complies with the Earth System Modeling Framework (ESMF) for portability across models. We present a new version here, HEMCO 3.0, that features an improved three-layer architecture to facilitate implementation into any atmospheric model and improved capability for calculating emissions at any model resolution including multiscale and unstructured grids. The three-layer architecture of HEMCO 3.0 includes (1) the Data Input Layer that reads the configuration file and accesses the HEMCO library of emission inventories and other environmental data, (2) the HEMCO Core that computes emissions on the user-selected HEMCO grid, and (3) the Model Interface Layer that re-grids (if needed) and serves the data to the atmospheric model and also serves model data to the HEMCO Core for computing emissions dependent on model state (such as from dust or vegetation). The HEMCO Core is common to the implementation in all models, while the Data Input Layer and the Model Interface Layer are adaptable to the model environment. Default versions of the Data Input Layer and Model Interface Layer enable straightforward implementation of HEMCO in any simple model architecture, and options are available to disable features such as re-gridding that may be done by independent couplers in more complex architectures. The HEMCO library of emission inventories and algorithms is continuously enriched through user contributions so that new inventories can be immediately shared across models. HEMCO can also serve as a general data broker for models to process input data not only for emissions but for any gridded environmental datasets. We describe existing implementations of HEMCO 3.0 in (1) the GEOS-Chem “Classic” chemical transport model with shared-memory infrastructure, (2) the high-performance GEOS-Chem (GCHP) model with distributed-memory architecture, (3) the NASA GEOS Earth System Model (GEOS ESM), (4) the Weather Research and Forecasting model with GEOS-Chem (WRF-GC), (5) the Community Earth System Model Version 2 (CESM2), and (6) the NOAA Global Ensemble Forecast System – Aerosols (GEFS-Aerosols), as well as the planned implementation in the NOAA Unified Forecast System (UFS). Implementation of HEMCO in CESM2 contributes to the Multi-Scale Infrastructure for Chemistry and Aerosols (MUSICA) by providing a common emissions infrastructure to support different simulations of atmospheric chemistry across scales.
This paper provides an implementation analysis of the linear discontinuous finite element method (LD-FEM) that spans the space of (l, x, y, z). A practical implementation of LD includes 1) selecting a computationally efficient algorithm to solve the 4 x 4 matrix system Ax = b that describes the angular flux in a mesh element, and 2) choosing how to store the data used to construct the matrix A and the vector b to either reduce memory consumption or increase computational speed. To analyze the first of these, three algorithms were selected to solve the 4 x 4 matrix equation: Cramer's rule, a streamlined implementation of Gaussian elimination, and LAPACK's Gaussian elimination subroutine dgesv. The results indicate that Cramer's rule and the streamlined Gaussian elimination algorithm perform nearly equivalently and outperform LAPACK's implementation of Gaussian elimination by a factor of 2. To analyze the second implementation detail, three formulations of the discretized LD-FEM equations were provided for implementation in a transport solver: 1) a low-memory formulation, which relies heavily on 'on-the-fly' calculations and less on the storage of pre-computed data, 2) a high-memory formulation, which pre-computes much of the data used to construct A and b, and 3) a reduced-memory formulation, which lies between the low - and high-memory formulations. These three formulations were assessed in the Jaguar transport solver based on relative memory footprint and computational speed for increasing mesh size and quadrature order. The results indicated that the memory savings of the low-memory formulation were not sufficient to warrant its implementation. The high-memory formulation resulted in a significant speed advantage over the reduced-memory option (10-50%), but also resulted in a proportional increase in memory consumption (5-45%) for increasing quadrature order and mesh count; therefore, the practitioner should weigh the system memory constraints against any required runtime constraints to determine which option is more suitable for his or her specific applications. (authors)
Despite its seemingly nonsensical cost, we show through modeling and simulation that redundant computation merits full consideration as a resilience strategy for next-generation systems. Without revolutionary breakthroughs in failure rates, part counts, or stable-storage bandwidths, it has been shown that the utility of Exascale systems will be crushed by the overheads of traditional checkpoint/restart mechanisms. Alternate resilience strategies must be considered, and redundancy is a proven unrivaled approach in many domains. We develop a distribution-independent model for job interrupts on systems of arbitrary redundancy, adapt Daly’s model for total application runtime, and find that his estimate for optimal checkpoint interval remains valid for redundant systems. We then identify conditions where redundancy is more cost effective than non-redundancy. These are done in the context of the number one supercomputers of the last decade, showing that thorough consideration of redundant computation is timely - if not overdue.
Systems, apparatuses, and methods for managing buffers in a neural network implementation with heterogeneous memory are disclosed. A system includes a neural network coupled to a first memory and a second memory. The first memory is a relatively low-capacity, high-bandwidth memory while the second memory is a relatively high-capacity, low-bandwidth memory. During a forward propagation pass of the neural network, a run-time manager monitors the usage of the buffers for the various layers of the neural network. During a backward propagation pass of the neural network, the run-time manager determines how to move the buffers between the first and second memories based on the monitored buffer usage during the forward propagation pass. As a result, the run-time manager is able to reduce memory access latency for the layers of the neural network during the backward propagation pass.
As part of the Consortium for Advanced Simulation of Light Water Reactors (CASL), the Virtual Environment for Reactor Applications (VERA) is being developed to provide high-fidelity multi-physics simulations of nuclear reactor cores. The MPACT code being developed collaboratively by Oak Ridge National Laboratory and the University of Michigan includes the primary deterministic neutron transport solver available within VERA. MPACT employs the 2D/1D method to solve three-dimensional problems, using 2D method of characteristics (MOC) radial transport solvers with 1D NEM-SP3 axial transport solvers, which are coupled using axial and radial transverse leakages. A 3D coarse mesh finite difference (CMFD) solver is used to provide accelerated convergence that is more stable. The Portable, Extensible Tool kit for Scientific Computation (PETSc) generalized minimal residual (GMRES) iterative solver with a block Jacobi preconditioner has been used to solve the CMFD linear systems. This work focused on the addition of a red-black successive over-relaxation (RBSOR) solver to reduce overall runtime by reducing time spent solving the CMFD linear systems. (authors)
We investigate the file transfer scheduling problem, where transfers among different endpoints must be scheduled to maximize pertinent metrics. We propose two new algorithms that exploit the fact that the aggregate bandwidth obtained over a network or at a storage system tends to increase with the number of concurrent transfers—but only up to a certain limit. The first algorithm, SEAL, uses runtime information and data-driven models to approximate system load and adapt transfer schedules and concurrency so as to maximize performance while avoiding saturation. We implement this algorithm using GridFTP as the transfer protocol and evaluate it using real transfer logs in a production WAN environment. Results show that SEAL can improve average slowdowns and turnaround times by up to 25% and worst-case slowdown and turnaround times by up to 50%, compared with the best-performing baseline scheme. Our second algorithm, STEAL, further leverages user-supplied categorization of transfers as either “interactive” (requiring immediate processing) or “batch” (less time-critical). Results show that STEAL reduces the average slowdown of interactive transfers by 63% compared to the best-performing baseline and by 21% compared to SEAL. For batch transfers, compared to the best-performing baseline, STEAL improves by 18% the utilization of the bandwidth unused by interactive transfers. By elegantly ensuring a sufficient, but not excessive, allocation of concurrency to the right transfers, we significantly improve overall performance despite constraints.
Modern large-scale scientific discovery requires multidisciplinary collaboration across diverse computing facilities, including High Performance Computing (HPC) machines and the Edge-to-Cloud continuum. Integrated data analysis plays a crucial role in scientific discovery, especially in the current AI era, by enabling Responsible AI development, FAIR, Reproducibility, and User Steering. However, the heterogeneous nature of science poses challenges such as dealing with multiple supporting tools, cross-facility environments, and efficient HPC execution. Building on data observability, adapter system design, and provenance, we propose MIDA: an approach for lightweight runtime Multi-workflow Integrated Data Analysis. MIDA defines data observability strategies and adaptability methods for various parallel systems and machine learning tools. With observability, it intercepts the dataflows in the background without requiring instrumentation while integrating domain, provenance, and telemetry data at runtime into a unified database ready for user steering queries. We conduct experiments showing end-to-end multi-workflow analysis integrating data from Dask and MLFlow in a real distributed deep learning use case for materials science that runs on multiple environments with up to 276 GPUs in parallel. We show near-zero overhead running up to 100,000 tasks on 1,680 CPU cores on the Summit supercomputer.