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.
We introduce the Binned Large Interval Network Kernel approach for scoring fragmentation mass spectra similarity which takes advantage of sparse matrix multiplication. This enables BLINK to perform 100 million comparisons while maintaining 98% accuracy in 2.2 seconds making it over 1,000 times faster than the current implementations.
This is the final technical report for the UNC-CH component of a SciDAC3 project. This project served to coordinate the software development effort which the nuclear physics lattice QCD community is developing to ensure that lattice calculations can make optimal use of forthcoming leadership-class and dedicated hardware, including those of the national laboratories, and prepares for the exploitation of future computational resources in the exascale era. The project developed and optimized simulation software for lattice QCD calculations on leadership class computers and on GPUaccelerated heterogeneous architectures. The participating team at multiple instritutions improved and extended software libraries (Chroma, CPS, MILC, QLUA, QPhiX, QDP++, QDP-jit) by providing interfaces for heterogeneous computing environments and through the optimization of gauge field evolution algorithms and sparse matrix inverters. The former includes the development of new multi-time scale integrators and the latter focused on the development of new multi-grid and domain decomposition inverters for several QCD fermion discretization schemes. The sub-project at the University of North Carolina worked towards the development a domain specific language for lattice QCD computations with the express purpose of generating code with uncompromised performance.
A parallel two-level overlapping Schwarz domain decomposition preconditioner has been integrated into the Trilinos ShyLU-package. The preconditioner uses an energy-minimizing coarse space and can be constructed from an assembled sparse matrix. The software implements variants of the two-level overlapping Schwarz method from [Dohrmann, Klawonn, Widlund, SINUM 2008], where it was denoted Generalized Dryja, Smith, Widlund (GDSW). The implementation is based on [Heinlein, Klawonn, Rheinbach, SISC 2016] but has been improved significantly with respect to efficiency, generality, e.g., for the use of Tpetra instead of Epetra matrices, and its interface.
This report describes a new capability for hierarchical task-data parallelism using Sandia's Kokkos and Qthreads, and evaluation of this capability with sparse matrix Cholesky factorization and social network triangle enumeration mini-applications. Hierarchical task-data parallelism consists of a collection of tasks with executes-after dependences where each task contains data parallel operations performed on a team of hardware threads. The collection of tasks and dependences form a directed acyclic graph of tasks - a task DAG. Major challenges of this research and development effort include: portability and performance across multicore CPU; manycore Intel Xeon Phi, and NVIDIA GPU architectures; scalability with respect to hardware concurrency and size of the task DAG; and usability of the application programmer interface (API).
Field-programmable gate arrays (FPGAs) are becoming more promising for high performance computing (HPC). Unlike fixed hardware specialization, reconfigurability allows FPGAs to be customized for a wide variety of HPC workloads. Our ultimate goal is to understand the viability of FPGAs as accelerators for some of the key HPC codes at LLNL and identify application characteristics that can benefit from the reconfigurability of FPGAs. More specifically, this LDRD project conducted a detailed performance study of FPGAs using sparse matrix computations and identified key optimization techniques to effectively use FPGA hardware resources.
Deployment and execution of machine learning tasks on extreme-scale computing platforms face several significant technical challenges: 1) High computing cost incurred by dense networks – The computing workload of deep networks with densely-connected topology increases rapidly with the network size, imposing a non-scalable computing model of extreme-scale computing platforms; 2) Non-optimized workload distribution – Many advanced deep learning algorithms, e.g., sparsification and irregular net-work topology, produce very unbalanced workload distribution on extreme-scale computing platforms. The computation efficiency is greatly hindered by the incurred data and computation redundancies as well as long tails of the node with extensive workload; 3) Constraints in data movement and I/O bottle-neck – Inter-node data movement in extreme-scale computing platforms are associated with high energy and latency costs, and subject to the constraints of I/O bandwidth; and 4) Generalization of algorithm realization and acceleration on computing platforms – The large varieties of machine learning algorithms and structures of extreme-scale computing platforms make the derivation of a generalized algorithm realization and acceleration method very challenging, which, however, is the requirement by domain scientists and interested users. We call the above challenges Smale’s Problems in Machine Learning and Understanding for High-Performance Computing Scientific Discovery. The objective of our three-year research project is to develop a holistic innovation set at structure, assembly, and acceleration layers of machine learning algorithms to address the above challenges in algorithm deployment and execution. Three tasks are particularly performed, including: At the algorithm structure level, we investigate the techniques that can structurally sparsify on the topology of deep networks for computing workload reduction. We also study clustering and pruning techniques that can optimize the workload distributions over the extreme-scale computing platforms; At the algorithm assembly level, we derive a unified learning framework for unsupervised transfer learning and dynamic growing capabilities. Novel training methods are also exploited to enhance the training efficiency of the proposed framework; At the algorithm acceleration level, we will develop a series of techniques that can accelerate the computation of sparse matrix operations, which are one of the core executions in deep learning and optimize memory access of the concerned platforms. Our proposed techniques attack the fundamental problems in machine learning algorithms running on extreme-scale computing platforms by vertically integrating the solutions at three closely entangled layers, paving the long-term scaling path of machine learning applications under DOE context. Three tasks corresponding to the above respective research orientations are performed during the three-year project period with our collaborators at ORNL. The outcome of the proposed project is anticipated to form a holistic solution set of novel algorithms and network topologies, efficient training techniques, and fast acceleration methods to promote the computing scalability of the machine learning applications of particular interest to DOE.
Lattice Quantumchromodynamics (QCD) is a powerful tool to numerically access the low energy regime of QCD in a straightforward way with quantifyable uncertainties. In this approach, QCD is discretized on a four dimensional, euclidian space-time grid with millions of degrees of freedom. In modern lattice calculations, most of the work is still spent in solving large, sparse linear systems. This part has two challenges, i.e. optimizing the sparse matrix application as well as BLAS-like kernels used in the linear solver. We are going to present performance optimizations of the Dirac operator (dslash) with and without clover term for recent Intel architectures, including Knights Landing (KNL). Using only a single right hand side, we were able to achieve a good fraction of peak performance for slash, CG and BiCGStab. We will also present a series of experiments we performed on KNL, i.e. running MCDRAM in different modes, enabling or disabling hardware prefetching, using different SoA lengths as well as different sub-NUMA-clustering modes.
The calculation of disconnected diagram contributions to physical signals is a computationally expensive task in Lattice QCD. To extract the physical signal, the trace of the inverse Lattice Dirac operator, a large sparse matrix, must be stochastically estimated. Because the variance of the stochastic estimator is typically large, variance reduction techniques must be employed. Multilevel Monte Carlo (MLMC) methods reduce the variance of the trace estimator by utilizing a telescoping sequence of estimators. Frequency Splitting is one such method that uses a sequence of inverses of shifted operators to estimate the trace of the inverse of the lattice Dirac operator, however there is no a priori way to select the shifts that minimize the cost of the multilevel trace estimation. We present a sampling and interpolation scheme that is able to predict the variances associated with Frequency Splitting under displacements of the underlying space time lattice. The interpolation scheme is able to predict the variances to high accuracy and therefore choose shifts that correspond to an approximate minimum of the cost for the trace estimation. We show that Frequency Splitting with the chosen shifts displays significant speedups over multigrid deflation
We present a quantum eigenstate filtering algorithm based on quantum signal processing (QSP) and minimax polynomials. The algorithm allows us to efficiently prepare a target eigenstate of a given Hamiltonian, if we have access to an initial state with non-trivial overlap with the target eigenstate and have a reasonable lower bound for the spectral gap. We apply this algorithm to the quantum linear system problem (QLSP), and present two algorithms based on quantum adiabatic computing (AQC) and quantum Zeno effect respectively. Both algorithms prepare the final solution as a pure state, and achieves the near optimal $\mathcal{\tilde{O}}$(dκlog(1/ϵ)) query complexity for a d -sparse matrix, where κ is the condition number, and ϵ is the desired precision. Neither algorithm uses phase estimation or amplitude amplification.
: As the rapid growth of marine renewable energy and off-shore wind energy, there have been concerns that the noises generated from construction and operation of the devices may interfere marine animals’ communication. In this research, a underwater sound model is developed to simulate sound prorogation generated by marine-hydrokinetic energy (MHK) devices or offshore wind (OSW) energy platforms. Finite volume and finite difference methods are developed to solve the 3D Helmholtz equation of sound propagation in the coastal environment. For finite volume method, the grid system consists of triangular grids in horizontal plane and sigma-layers in vertical dimension. A 3D sparse matrix solver with complex coefficients is formed for solving the resulting acoustic pressure field. The Complex Shifted Laplacian Preconditioner (CSLP) method is applied to efficiently solve the matrix system iteratively with MPI parallelization using a high performance cluster. The sound model is then coupled with the Finite Volume Community Ocean Model (FVCOM) for simulating sound propagation generated by human activities in a range-dependent setting, such as offshore wind energy platform constructions and tidal stream turbines. As a proof of concept, initial validation of the finite difference solver is presented for two coastal wedge problems. Validation of finite volume method will be reported separately.
An underwater sound model was developed to simulate sound propagation from marine and hydrokinetic energy (MHK) devices or offshore wind (OSW) energy platforms. Finite difference methods were developed to solve the 3D Helmholtz equation for sound propagation in the coastal environment. A 3D sparse matrix solver with complex coefficients was formed for solving the resulting acoustic pressure field. The Complex Shifted Laplacian Preconditioner (CSLP) method was applied to solve the matrix system iteratively with MPI parallelization using a high performance cluster. The sound model was then coupled with the Finite Volume Community Ocean Model (FVCOM) for simulating sound propagation generated by human activities, such as construction of OSW turbines or tidal stream turbine operations, in a range-dependent setting. As a proof of concept, initial validation of the solver is presented for two coastal wedge problems. This sound model can be useful for evaluating impacts on marine mammals due to deployment of MHK devices and OSW energy platforms.
Policies to mitigate greenhouse gas (GHG) emissions will not only slow climate change but can also have ancillary benefits of improved air quality. Here we examine the co-benefits of both global and regional GHG mitigation for US air quality in 2050 at fine resolution, using dynamical downscaling methods, building on a previous global co-benefits study (West et al., 2013). The co-benefits for US air quality are quantified via two mechanisms: through reductions in co-emitted air pollutants from the same sources and by slowing climate change and its influence on air quality, following West et al. (2013). Additionally, we separate the total co-benefits into contributions from domestic GHG mitigation vs. mitigation in foreign countries. We use the Weather Research and Forecasting (WRF) model to dynamically downscale future global climate to the regional scale and the Sparse Matrix Operator Kernel Emissions (SMOKE) program to directly process global anthropogenic emissions to the regional domain, and we provide dynamical boundary conditions from global simulations to the regional Community Multi-scale Air Quality (CMAQ) model. The total co-benefits of global GHG mitigation from the RCP4.5 scenario compared with its reference are estimated to be higher in the eastern US (ranging from 0.6 to 1.0 µg m -3 ) than the west (0–0.4 µg m -3 ) for fine particulate matter (PM 2.5 ), with an average of 0.47 µg m -3 over the US; for O 3 , the total co-benefits are more uniform at 2–5 ppb, with a US average of 3.55 ppb. Comparing the two mechanisms of co-benefits, we find that reductions in co-emitted air pollutants have a much greater influence on both PM 2.5 (96 % of the total co-benefits) and O 3 (89 % of the total) than the second co-benefits mechanism via slowing climate change, consistent with West et al. (2013). GHG mitigation from foreign countries contributes more to the US O 3 reduction (76 % of the total) than that from domestic GHG mitigation only (24 %), highlighting the importance of global methane reductions and the intercontinental transport of air pollutants. For PM 2.5 , the benefits of domestic GHG control are greater (74 % of total). Since foreign contributions to co-benefits can be substantial, with foreign O 3 benefits much larger than those from domestic reductions, previous studies that focus on local or regional co-benefits may greatly underestimate the total co-benefits of global GHG reductions. We conclude that the US can gain significantly greater domestic air quality co-benefits by engaging with other nations to control GHGs.
We investigate the use of sparse matrices and OpenMP multi-threading on linear algebra operations involving them. Several sparse matrix data structures are presented. Implementation of the multi- threading primarily occurs in the level one and two BLAS functions used within the four algorithms investigated{the Power Method, Conjugate Gradient, Biconjugate Gradient, and Jacobi's Method. The bene ts of launching threads once per high level algorithm are explored.
Roughness-induced transient disturbances distort laminar boundary layers and can lead to premature transition if the roughness is of sufficient height and shape. If transition does occur, it does so through a high-frequency "secondary" instability of the roughness wake. A BiGlobal secondaryinstability code has been developed to solve the stability problem downstream of roughness. The code implements multiple solvers for incompressible/compressible, inviscid/viscous, and temporal/spatial stability. High-order finite differences, sparse matrix linear algebra, LU decomposition, and Arnoldi iterations are used to solve the resulting generalized eigenvalue problem. The code has been used to evaluate the stability properties of experimental roughness-induced transient growth on a flat-plate at low-speed and a slightly blunted cone at M∞= 6. The resulting mode shapes, growth rates, and spatial N-factors show excellent agreement with experimental data. The viscous mode shapes show the best agreement with the experimental disturbance profiles. Yet, execution time can be reduced by an order-of-magnitude by using the inviscid solvers with only modest differences in growth rates.
Sparse neural network attracts more attention for model compression, fast execution, and power reduction. The state-of-the-art designed sparse tensor core for structured and static sparsity, which did not support well for generic or dynamic sparsity. We design a sparse tensor core to support generic sparsity pruning with a novel hybrid and blocked sparse matrix storage format, HB-ELL, which saves computation and storage while keeping the most significant elements, as well as supporting dynamic sparsity for data flow in neural networks. We achieve better performance with preliminary results than the state-of- the-art on an NVIDIA GPU simulator.
Direct Numerical Simulations of realistic combustion devices are extremely challenging due to the wide separation of scales in the simulation, for example an internal combustion (IC) engine chamber, and the flame thickness of a high-pressure flame. The PeleLMeX solver uses adaptive mesh refinement (AMR) to evolve multi-species reacting flows in the low Mach number limit at the Exascale and relies on an embedded boundary (EB) approach to represent complex geometries. In that framework, the EB geometries often give rise to very small cut-cells along the boundary, which translate into extreme ill-conditioning of the pressure-projection, with eigenvalues that span 15-16 orders of magnitude. In this talk, we focus on the case of a typical IC piston bowl geometry for which we present on a novel approach towards solving these nearly singular linear systems with ILU-based, C-AMG smoothers on massively parallel architectures. In particular, we use scaling and equilibration algorithms to handle the non-normality of the upper triangular factors. This enables us to approximate the highly sequential triangular solve algorithm, embedded in the AMG smoothing-solve phase, with Jacobi iterations. This approximation can be written as a convergent Neumann series whose terms are composed of highly parallel sparse matrix vector multiplications. The result is an algorithm that substantially decreases setup and solve time, compared to state-of-the-art, for these challenging linear systems.
We describe the effort to implement the HPCG benchmark using OpenSHMEM and MPI one-sided communication. Unlike the High Performance LINPACK (HPL) benchmark that places emphasis on large dense matrix computations, the HPCG benchmark is dominated by sparse operations such as sparse matrix-vector product, sparse matrix triangular solve, and long vector operations. The MPI one-sided implementation is developed using the one-sided OpenSHMEM implementation. Preliminary results comparing the original MPI, OpenSHMEM, and MPI one-sided implementations on an SGI cluster, Cray XK7 and Cray XC30 are presented. The results suggest the MPI, OpenSHMEM, and MPI one-sided implementations all obtain similar overall performance but the MPI one-sided implementation seems to slightly increase the run time for multigrid preconditioning in HPCG on the Cray XK7 and Cray XC30.
We describe the effort to implement the HPCG benchmark using OpenSHMEM and MPI one-sided communication. Unlike the High Performance LINPACK (HPL) benchmark that places em- phasis on large dense matrix computations, the HPCG benchmark is dominated by sparse operations such as sparse matrix-vector product, sparse matrix triangular solve, and long vector operations. The MPI one-sided implementation is developed using the one-sided OpenSHMEM implementation. Pre- liminary results comparing the original MPI, OpenSHMEM, and MPI one-sided implementations on an SGI cluster, Cray XK7 and Cray XC30 are presented. The results suggest the MPI, OpenSHMEM, and MPI one-sided implementations all obtain similar overall performance but the MPI one-sided im- plementation seems to slightly increase the run time for multigrid preconditioning in HPCG on the Cray XK7 and Cray XC30.