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.
We present performance results for version 2.1 of the NAS Parallel Benchmarks (NPB) on the following architectures: IBM SP2/66 MHz; SGI Power Challenge Array/90 MHz; Cray Research T3D; and Intel Paragon. The NAS Parallel Benchmarks are a widely-recognized suite of benchmarks originally designed to compare the performance of highly parallel computers with that of traditional supercomputers.
The implementation and performance of a parallel spatial direct numerical simulation (PSDNS) approach on the Intel iPSC/860 hypercube is documented. The direct numerical simulation approach is used to compute spatially evolving disturbances associated with the laminar-to-turbulent transition in boundary-layer flows. The feasibility of using the PSDNS on the hypercube to perform transition studies is examined. The results indicate that the direct numerical simulation approach can effectively be parallelized on a distributed-memory parallel machine. By increasing the number of processors nearly ideal linear speedups are achieved with nonoptimized routines; slower than linear speedups are achieved with optimized (machine dependent library) routines. This slower than linear speedup results because the Fast Fourier Transform (FFT) routine dominates the computational cost and because the routine indicates less than ideal speedups. However with the machine-dependent routines the total computational cost decreases by a factor of 4 to 5 compared with standard FORTRAN routines. The computational cost increases linearly with spanwise wall-normal and streamwise grid refinements. The hypercube with 32 processors was estimated to require approximately twice the amount of Cray supercomputer single processor time to complete a comparable simulation; however it is estimated that a subgrid-scale model which reduces the required number of grid points and becomes a large-eddy simulation (PSLES) would reduce the computational cost and memory requirements by a factor of 10 over the PSDNS. This PSLES implementation would enable transition simulations on the hypercube at a reasonable computational cost.
A method is presented for executing a direct simulation Monte Carlo (DSMC) analysis using parallel processing. The method is based on using domain decomposition to distribute the work load among multiple processors, and the DSMC analysis is performed completely in parallel. Message passing is used to transfer molecules between processors and to provide the synchronization necessary for the correct physical simulation. Benchmark problems are described for testing the method and results are presented which demonstrate the performance on two commercially available multicomputers. The results show that reasonable parallel speedup and efficiency can be obtained if the problem is properly sized to the number of processors. It is projected that with a massively parallel system, performance exceeding that of current supercomputers is possible.
Irregular communication often limits both the performance and scalability of parallel applications. Typically, applications individually implement irregular communication as point-to-point, and any optimizations are integrated directly into the application. As a result, these optimizations lack portability. It is difficult to optimize point-to-point messages within MPI, as the interface for single messages provides no information on the collection of all communication to be performed. However, the persistent neighbor collective API, released in the MPI 4 standard, provides an interface for portable optimizations of irregular communication within MPI libraries. This paper presents methods for implementing existing optimizations for irregular communication within neighborhood collectives, analyzes the impact of replacing point-to-point communication in existing codebases such as Hypre BoomerAMG with neighborhood collectives, and finally shows up to a 1.38x speedup on sparse matrix-vector multiplication communication within a BoomerAMG solve through the use of our optimized neighbor collectives. Here, the authors analyze three implementations of persistent neighborhood collectives for Alltoallv: an unoptimized wrapper of standard point-to-point communication, and two locality-aware aggregating methods. The second locality-aware implementation exposes an non-standard interface to perform additional optimization, and the authors present the additional 0.07x speedup from the extended interface. All optimizations are available in an open-source codebase, MPI Advance, which sits on top of MPI, allowing for optimizations to be added into existing codebases regardless of the system MPI install.
(view graphs) Beowulf System (Hyglac), Hyglac cost, Hyglac Performance vs. T3D, FDTD Interior Communication, Boundary Communication, Timing.
In this work, we collect data from runs of Krylov subspace methods and pipelined Krylov algorithms in an effort to understand and model the impact of machine noise and other sources of variability on performance. We find large variability of Krylov iterations between compute nodes for standard methods that is reduced in pipelined algorithms, directly supporting conjecture, as well as large variation between statistical distributions of runtimes across iterations. Based on these results, we improve upon a previously introduced nondeterministic performance model by allowing iterations to fluctuate over time. We present our data from runs of various Krylov algorithms across multiple platforms as well as our updated non-stationary model that provides good agreement with observations. We also suggest how it can be used as a predictive tool.
Ames Instrumentation System (AIMS) computer program package of software tools measuring and analyzing performances of parallel-processing application programs. Helps programmer to debug and refine, and to monitor and visualize execution of, parallel-processing application software for Intel iPSC/860 (or equivalent) multicomputer. Performance data collected displayed graphically on computer workstations supporting X-Windows.
Improvements and advances in the development of computer architecture now provide innovative technology for the recasting of traditional sequential solutions into high-performance, low-cost, parallel system to increase system performance. Research conducted in development of specialized computer architecture for the algorithmic execution of an avionics system, guidance and control problem in real time is described. A comprehensive treatment of both the hardware and software structures of a customized computer which performs real-time computation of guidance commands with updated estimates of target motion and time-to-go is presented. An optimal, real-time allocation algorithm was developed which maps the algorithmic tasks onto the processing elements. This allocation is based on the critical path analysis. The final stage is the design and development of the hardware structures suitable for the efficient execution of the allocated task graph. The processing element is designed for rapid execution of the allocated tasks. Fault tolerance is a key feature of the overall architecture. Parallel numerical integration techniques, tasks definitions, and allocation algorithms are discussed. The parallel implementation is analytically verified and the experimental results are presented. The design of the data-driven computer architecture, customized for the execution of the particular algorithm, is discussed.
Recently, a Cray T3D Emulator has been made available on the Cray Y-MP and C90 computers. The Pittsburgh Supercomputer Center has acquired a CRAY T3D system and many other centers like Jet Propulsion Laboratory (JPL) will have it by the end of 1994. The Cray T3D system is the firstphase system in Cray Research, Inc.'s (CRI) three-phase massively parallel processing (MPP) program. This system features a heterogeneous architecture that closely couples DEC's ALPHA microprocessors and CRI's parallel-vector technology, i.e. the Cray Y-MP and Cray C90. The Cray T3D Emulator will give prospective users a valuable experience in developing high performance applications on the MPP system. This emulator runs programs written in CRI's MPP Fortran programming model (data sharing and work sharing) or Parallel Virtual Machine (PVM) programming model. It will help the users to study data layout, data locality, and data reference patterns thereby providing feedback which will enable one to write more efficient parallel codes. An overview of the Cray T3D hardware, software, and three of its available programming models is presented.The Cray Fortran Programming Model comprising (a) Data Sharing, (b) Worksharing and (c) Message Passing, will be discussed with examples. We have also implemented distributed BLAS 3 (matrix-matrix multiplication) in data parallel model (using only CSHIFT); worksharing model using block distribution and collapsed distribution; and message passing model using PVM. We have also implemented 2D and 3D FFTs for radix-2 using PVM. The performance of NAS Parallel 'Benchmarks (NPB) on CRAY T3D will be compared with other highly parallel systems such as CM-5, Paragon, C90 etc.
The implementation and performance of a parallel spatial direct numerical simulation (PSDNS) approach on the Intel iPSC/860 hypercube and IBM SP1 and SP2 parallel computers is documented. Spatially evolving disturbances associated with the laminar-to-turbulent transition in boundary-layer flows are computed with the PSDNS code. The feasibility of using the PSDNS to perform transition studies on these computers is examined. The results indicate that PSDNS approach can effectively be parallelized on a distributed-memory parallel machine by remapping the distributed data structure during the course of the calculation. Scalability information is provided to estimate computational costs to match the actual costs relative to changes in the number of grid points. By increasing the number of processors, slower than linear speedups are achieved with optimized (machine-dependent library) routines. This slower than linear speedup results because the computational cost is dominated by FFT routine, which yields less than ideal speedups. By using appropriate compile options and optimized library routines on the SP1, the serial code achieves 52-56 M ops on a single node of the SP1 (45 percent of theoretical peak performance). The actual performance of the PSDNS code on the SP1 is evaluated with a "real world" simulation that consists of 1.7 million grid points. One time step of this simulation is calculated on eight nodes of the SP1 in the same time as required by a Cray Y/MP supercomputer. For the same simulation, 32-nodes of the SP1 and SP2 are required to reach the performance of a Cray C-90. A 32 node SP1 (SP2) configuration is 2.9 (4.6) times faster than a Cray Y/MP for this simulation, while the hypercube is roughly 2 times slower than the Y/MP for this application. KEY WORDS: Spatial direct numerical simulations; incompressible viscous flows; spectral methods; finite differences; parallel computing.
Abstract Machine learning (ML) continues to grow in importance across nearly all domains in modeling to learn from data. Often a tradeoff exists between a model's ability to minimize bias and variance. In this article, we utilize ensemble learning to combine linear, nonlinear, and tree‐/rule‐based ML methods to cope with the bias‐variance tradeoff and result in more accurate models. We use the datasets collected for two parallel cancer deep learning CANDLE benchmarks, NT3 and P1B2, to build performance and power models based on hardware performance counters using single‐object and multiple‐objects ensemble learning to identify the most important counters for improvement on the Cray XC40 Theta at Argonne National Laboratory. Based on the insights from these models, we improve the performance and energy of P1B2 and NT3 by optimizing the deep learning environments TensorFlow, Keras, Horovod, and Python under the huge page size of 8 MB. Experimental results show that ensemble learning not only produces more accurate models but also provides more robust performance counter ranking. We achieve up to 61.15% performance improvement and up to 62.58% energy saving for P1B2 and up to 55.81% performance improvement and up to 52.60% energy saving for NT3 on up to 24,576 cores.
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.
The purpose of this study is to investigate the performance of the integral equation computations using numerical source field-panel method in a massively parallel processing (MPP) environment. A comparative study of computational performance of the MPP CM-5 computer and conventional Cray-YMP supercomputer for a three-dimensional flow problem is made. A serial FORTRAN code is converted into a parallel CM-FORTRAN code. Some performance results are obtained on CM-5 with 32, 62, 128 nodes along with those on Cray-YMP with a single processor. The comparison of the performance indicates that the parallel CM-FORTRAN code near or out-performs the equivalent serial FORTRAN code for some cases.
Current plans for independent instrument approaches to closely spaced parallel runways call for an automated pilot alerting system to ensure separation of aircraft in the case of a "blunder," or unexpected deviation from the a normal approach path. Resolution advisories by this system would require the pilot of an endangered aircraft to perform a trained evasion maneuver. The potential performance of two evasion maneuvers, referred to as the "turn-climb" and "climb-only," was estimated using an experimental NASA alerting logic (AILS) and a computer simulation of relative trajectory scenarios between two aircraft. One aircraft was equipped with the NASA alerting system, and maneuvered accordingly. Observation of the rates of different types of alerting failure allowed judgement of evasion maneuver performance. System Operating Characteristic (SOC) curves were used to assess the benefit of alerting with each maneuver.
We perform a 2.5-dimensional particle-in-cell simulation of a quasi-parallel shock, using parameters for the Earth's bow shock, to examine electron acceleration and heating due to magnetic reconnection. The shock transition region evolves from the ion-coupled reconnection dominant stage to the electron-only reconnection dominant stage, as time elapses. The electron temperature enhances locally in each reconnection site, and ion-scale magnetic islands generated by ion-coupled reconnection show the most significant enhancement of the electron temperature. The electron energy spectrum shows a power law, with a power-law index around 6. We perform electron trajectory tracing to understand how they are energized. Some electrons interact with multiple electron-only reconnection sties, and Fermi acceleration occurs during multiple reflections. Electrons trapped in ion-scale magnetic islands can be accelerated in another mechanism. Islands move in the shock transition region, and electrons can obtain larger energy from the in-plane electric field than the electric potential in those islands. These newly found energization mechanisms in magnetic islands in the shock can accelerate electrons to energies larger than the achievable energies by the conventional energization due to the parallel electric field and shock drift acceleration. This study based on the selected particle analysis indicates that the maximum energy in the nonthermal electrons is achieved through acceleration in ion-scale islands, and electron-only reconnection accounts for no more than half of the maximum energy, as the lifetime of sub-ion-scale islands produced by electron-only reconnection is several times shorter than that of ion-scale islands.
Whether a researcher is designing the 'next parallel programming paradigm', another 'scalable multiprocessor' or investigating resource allocation algorithms for multiprocessors, a facility that enables parallel program execution to be captured and displayed is invaluable. Careful analysis of such information can help computer and software architects to capture, and therefore, exploit behavioral variations among/within various parallel programs to take advantage of specific hardware characteristics. A software tool-set that facilitates performance evaluation of parallel applications on multiprocessors has been put together at NASA Ames Research Center under the sponsorship of NASA's High Performance Computing and Communications Program over the past five years. The Automated Instrumentation and Monitoring Systematic has three major software components: a source code instrumentor which automatically inserts active event recorders into program source code before compilation; a run-time performance monitoring library which collects performance data; and a visualization tool-set which reconstructs program execution based on the data collected. Besides being used as a prototype for developing new techniques for instrumenting, monitoring and presenting parallel program execution, AIMS is also being incorporated into the run-time environments of various hardware testbeds to evaluate their impact on user productivity. Currently, the execution of FORTRAN and C programs on the Intel Paragon and PALM workstations can be automatically instrumented and monitored. Performance data thus collected can be displayed graphically on various workstations. The process of performance tuning with AIMS will be illustrated using various NAB Parallel Benchmarks. This report includes a description of the internal architecture of AIMS and a listing of the source code.
Over the past decade, high performance computing has evolved rapidly; systems based on commodity microprocessors have been introduced in quick succession from at least seven vendors/families. Porting codes to every new architecture is a difficult problem; in particular, here at NASA, there are many large CFD applications that are very costly to port to new machines by hand. The LCM ("Legacy Code Modernization") Project is the development of an integrated parallelization environment (IPE) which performs the automated mapping of legacy CFD (Fortran) applications to state-of-the-art high performance computers. While most projects to port codes focus on the parallelization of the code, we consider porting to be an iterative process consisting of several steps: 1) code cleanup, 2) serial optimization,3) parallelization, 4) performance monitoring and visualization, 5) intelligent tools for automated tuning using performance prediction and 6) machine specific optimization. The approach for building this parallelization environment is to build the components for each of the steps simultaneously and then integrate them together. The demonstration will exhibit our latest research in building this environment: 1. Parallelizing tools and compiler evaluation. 2. Code cleanup and serial optimization using automated scripts 3. Development of a code generator for performance prediction 4. Automated partitioning 5. Automated insertion of directives. These demonstrations will exhibit the effectiveness of an automated approach for all the steps involved with porting and tuning a legacy code application for a new architecture.
Simulation of multiphase flow in the subsurface is well-known to be computationally challenging. While there have been many studies that have explored approaches to overcoming these challenges, they often utilize relatively simple case studies. In this paper, we focus on the unique numerical challenges posed by modeling large-scale engineered subsurface systems, characterized by discrete features embedded in a heterogeneous natural subsurface setting. The man-made features such as shafts, tunnels, and barriers often cause multiple challenges in modeling the domain for multiphase porous media flow. This flow scenario can have a wide range of applications such as nuclear waste repositories, enhanced recovery of a petroleum reservoir, geothermal engineering, and carbon sequestration. An example of these severe numerical challenges is the case of performance assessment (PA) for Waste Isolation Pilot Plant (WIPP), the only operating deep geological repository in the US, which simulates extreme material properties of bedded salt rock formation and extreme contrast due to open excavation next to the formation. The models have extremes not only of permeability and porosity but also of the constitutive models needed for multiphase flow; additionally, they have process models like salt creep closure reducing porosity over time, fracturing in clay and anhydrite interbeds of the bedded salt, gas generation from the waste materials, and unintentional human borehole intrusions in some scenarios. Numerical simulations require the solution of coupled systems of nonlinear PDEs; in our work, we use the open-source simulator PFLOTRAN which is based on Finite Volume discretization. The solution of the nonlinear equations requires use of the Newton-Raphson iteration at each time step, which entails the solution of the linearized Jacobian system at each iteration. The effects of all the processes (i.e., large number of unknowns, highly nonlinear constitutive relations, large contrasts in material properties in short distances) lead to an ill-conditioned Jacobian matrix that severely challenges traditional linear solver, i.e., stabilized biconjugate gradient with block Jacobi incomplete LU preconditioner (BCGS-ILU) leading to non-convergence for traditional Newton-Raphson nonlinear solver causing unacceptably long computation time for each model. This paper presents linear solvers such as constrained pressure residual (CPR) two-stage preconditioner with alternate-block-factorization (ABF) and quasi- implicit pressure and explicit saturation (QIMPES) decouplers and flexible generalized residual solver (FGMRES). The new general-purpose nonlinear solver, Newton trust-region dogleg Cauchy (NTRDC), is also introduced to resolve extreme nonlinearities in the models. We demonstrate the effectiveness of each method relative to the default BCGS-Newton solver. The two best cases had nearly 50 times speed-up and achieved completion of a simulation in 14 hours that never completed due to non-convergence with the default solver. We also investigate the strong scalability of each method and discuss some of the deficiencies found for Block Jacobi preconditioner using parallel domain decomposition, and node packing effects of modern processor architecture.