Application Acceleration on Current and Future Cray Platforms
Abstract not provided
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.
Abstract not provided
Incompressible viscous turbulent flows over single- and multiple-element airfoils are numerically simulated in an efficient manner by solving the incompressible Navier-Stokes equations. The solution algorithm uses the method of pseudocompressibility with an upwind-differencing scheme for the convective fluxes and an implicit line-relaxation scheme to study high-lift take-off and landing configurations and to compute lift and drag at various angles of attack up to stall. Two different turbulence models are tested in computing the flow over an NACA 4412 airfoil. The approach used for multiple-element airfoils involves the use of multiple zones of structured grids fitted to each element. Two different approaches are compared: a patched system of grids and an overlaid Chimera system of grids. Computational results are presented for two-element, three-element, and four-element airfoil configurations. Excellent agreement with experimental surface-pressure coefficients is seen. The code converges in less than 200 iterations, requiring on the order of one minute of CPU time on a CRAY YMP per element in the airfoil configuration.
The theory and practice of implementing Tanemura's algorithm for 3D Delaunay triangulation on Intel's Gamma prototype, a 128 processor MIMD computer, is described. Efficient implementation of Tanemura's algorithm on a conventional, vector processing supercomputer is problematic. It does not vectorize to any significant degree and requires indirect addressing. Efficient implementation on a parallel architecture is possible, however. Speeds in excess of 20 times a single processor Cray Y-MP are realized on 128 processors of the Intel Gamma prototype.
Reactive molecular dynamics simulations are computationally demanding. Reaching spatial and temporal scales where interesting scientific phenomena can be observed requires efficient and scalable implementations on modern hardware. In this article, we focus on optimizing the performance of the widely used LAMMPS/ReaxC package for many-core architectures. As hybrid parallelism allows better leverage of the increasing on-node parallelism, we adopt thread parallelism in the construction of bonded and nonbonded lists and in the computation of complex ReaxFF interactions. To mitigate the I/O overheads due to large volumes of trajectory data produced and to save users the burden of post-processing, we also develop a novel in situ tool for molecular species analysis. We analyze the performance of the resulting ReaxC-OMP package on two different architectures: (i) Mira, an IBM Blue Gene/Q system and (ii) Cori-II, a Cray XC-40 sytem with Knights Landing processors. For Pentaerythritol tetranitrate (PETN) systems of sizes ranging from 32 thousand to 16.6 million particles, we observe speedups in the range of 1.5-4.5x. We observe sustained performance improvements for up to 262,144 cores (1,048,576 processes) of Mira and a weak scaling efficiency of 91.5% in large simulations containing 16.6 million particles. Finally, the in situ molecular species analysis tool incurs only insignificant overheads across various system sizes and runs configurations.
Although logically available, applications may not exploit enough instantaneous communication concurrency to maximize hardware utilization on HPC systems. This is exacerbated in hybrid programming models such as SPMD+OpenMP. We present the design of a "multi-threaded" runtime able to transparently increase the instantaneous network concurrency and to provide near saturation bandwidth, independent of the application configuration and dynamic behavior. The runtime forwards communication requests from application level tasks to multiple communication servers. Our techniques alleviate the need for spatial and temporal application level message concurrency optimizations. Experimental results show improved message throughput and bandwidth by as much as 150% for 4KB bytes messages on InfiniBand and by as much as 120% for 4KB byte messages on Cray Aries. For more complex operations such as all-to-all collectives, we observe as much as 30% speedup. This translates into 23% speedup on 12,288 cores for a NAS FT implemented using FFTW. We also observe as much as 76% speedup on 1,500 cores for an already optimized UPC+OpenMP geometric multigrid application using hybrid parallelism.
Deep Learning (DL) algorithms have become ubiquitous in data analytics. As a result, major computing vendors—including NVIDIA, Intel, AMD, and IBM—have architectural road maps influenced by DL workloads. Furthermore, several vendors have recently advertised new computing products as accelerating large DL workloads. Unfortunately, it is difficult for data scientists to quantify the potential of these different products. Here, this article provides a performance and power analysis of important DL workloads on two major parallel architectures: NVIDIA DGX-1 (eight Pascal P100 GPUs interconnected with NVLink) and Intel Knights Landing (KNL) CPUs interconnected with Intel Omni-Path or Cray Aries. Our evaluation consists of a cross section of convolutional neural net workloads: CifarNet, AlexNet, GoogLeNet, and ResNet50 topologies using the Cifar10 and ImageNet datasets. The workloads are vendor-optimized for each architecture. We use sequentially equivalent implementations to maintain iso-accuracy between parallel and sequential DL models. Our analysis indicates that although GPUs provide the highest overall performance, the gap can close for some convolutional networks; and the KNL can be competitive in performance/watt. We find that NVLink facilitates scaling efficiency on GPUs. However, its importance is heavily dependent on neural network architecture. Furthermore, for weak-scaling—sometimes encouraged by restricted GPU memory—NVLink is less important.
Collective communication operations, used by many scientific applications, tend to limit overall parallel application performance and scalability. Computer systems are becoming more heterogeneous with increasing node and core-per-node counts. Also, a growing number of data-access mechanisms, of varying characteristics, are supported within a single computer system. We describe a new hierarchical collective communication framework that takes advantage of hardware-specific data-access mechanisms. It is flexible, with run-time hierarchy specification, and sharing of collective communication primitives between collective algorithms. Data buffers are shared between levels in the hierarchy reducing collective communication management overhead. We have implemented several versions of the Message Passing Interface (MPI) collective operations, MPI Barrier() and MPI Bcast(), and run experiments using up to 49, 152 processes on a Cray XT5, and a small InfiniBand based cluster. At 49, 152 processes our barrier implementation outperforms the optimized native implementation by 75%. 32 Byte and one Mega-Byte broadcasts outperform it by 62% and 11%, respectively, with better scalability characteristics. Improvements relative to the default Open MPI implementation are much larger.
The human nervous system is a massively parallel processor of information. The vast numbers of neurons, synapses and circuits is daunting to those seeking to understand the neural basis of consciousness and intellect. Pervading obstacles are lack of knowledge of the detailed, three-dimensional (3-D) organization of even a simple neural system and the paucity of large scale, biologically relevant computer simulations. We use high performance graphics workstations and supercomputers to study the 3-D organization of gravity sensors as a prototype architecture foreshadowing more complex systems. Scaled-down simulations run on a Silicon Graphics workstation and scale-up, three-dimensional versions run on the Cray Y-MP and CM5 supercomputers.
We present a new method for mapping applications' MPI tasks to cores of a parallel computer such that applications' communication time is reduced. We address the case of sparse node allocation, where the nodes assigned to a job are not necessarily located in a contiguous block nor within close proximity to each other in the network, although our methods generalize to contiguous allocations as well. The goal is to assign tasks to cores so that interdependent tasks are performed by "nearby' cores, thus lowering the distance messages must travel, the amount of congestion in the network, and the overall cost of communication. Our new method applies a geometric partitioning algorithm to both the tasks and the processors, and assigns task parts to the corresponding processor parts. We also present a number of algorithmic optimizations that exploit specific features of the network or application. We show that, for the structured finite difference mini-application MiniGhost, our mapping methods reduced communication time up to 75% relative to MiniGhost's default mapping on 128K cores of a Cray XK7 with sparse allocation. For the atmospheric modeling code E3SM/HOMME, our methods reduced communication time up to 31% on 32K cores of an IBM BlueGene/Q with contiguous allocation.
A class of implicit upwind-differencing methods for the compressible Navier-Stokes equations is described and applied. The methods are based on the use of local eigenvalues or wave speeds to control spatial differencing of inviscid terms and are aimed at increasing the level of accuracy and stability achievable in computation. Techniques for accelerating the rate of convergence to a steady-state solution are also used. Applications to inviscid and viscous transonic flows are discussed and compared with other methods and experimental measurements. It is shown that accurate and efficient transonic airfoil calculations can be made on the Cray-1 coomputer in less than 2 min.
Progress in the development of computational methods for the characterization of chemical reactions in aerobraking orbit transfer vehicle (AOTV) propulsive flows is reported. Two main areas of code development were undertaken: (1) the implementation of CASSCF (complete active space self-consistent field) and SCF (self-consistent field) analytical first derivatives on the CRAY X-MP; and (2) the installation of the complete set of electronic structure codes on the CRAY 2. In the area of application calculations the main effort was devoted to performing full configuration-interaction calculations and using these results to benchmark other methods. Preprints describing some of the systems studied are included.
During the past two decades, there has been significant progress in the field of numerical simulation of unsteady compressible viscous flows. At present, a variety of solution techniques exist such as the transonic small disturbance analyses (TSD), transonic full potential equation-based methods, unsteady Euler solvers, and unsteady Navier-Stokes solvers. These advances have been made possible by developments in three areas: (1) improved numerical algorithms; (2) automation of body-fitted grid generation schemes; and (3) advanced computer architectures with vector processing and massively parallel processing features. In this work, the GMRES scheme has been considered as a candidate for acceleration of a Newton iteration time marching scheme for unsteady 2-D and 3-D compressible viscous flow calculation; from preliminary calculations, this will provide up to a 65 percent reduction in the computer time requirements over the existing class of explicit and implicit time marching schemes. The proposed method has ben tested on structured grids, but is flexible enough for extension to unstructured grids. The described scheme has been tested only on the current generation of vector processor architecture of the Cray Y/MP class, but should be suitable for adaptation to massively parallel machines.
This paper presents an architecture template for next-generation high performance computing systems specifically targeted to irregular applications. We start our work by considering that future generation interconnection and memory bandwidth full-system numbers are expected to grow by a factor of 10. In order to keep up with such a communication capacity, while still resorting to fine-grained multithreading as the main way to tolerate unpredictable memory access latencies of irregular applications, we show how overall performance scaling can benefit from the multi-core paradigm. At the same time, we also show how such an architecture template must be coupled with specific techniques in order to optimize bandwidth utilization and achieve the maximum scalability. We propose a technique based on memory references aggregation, together with the related hardware implementation, as one of such optimization techniques. We explore the proposed architecture template by focusing on the Cray XMT architecture and, using a dedicated simulation infrastructure, validate the performance of our template with two typical irregular applications. Our experimental results prove the benefits provided by both the multi-core approach and the bandwidth optimization reference aggregation technique.
We present performance results and an analysis of a message passing interface (MPI)/OpenACC implementation of an electromagnetic solver based on a spectral-element discontinuous Galerkin discretization of the time-dependent Maxwell equations. The OpenACC implementation covers all solution routines, including a highly tuned element-by-element operator evaluation and a GPUDirect gather–scatter kernel to effect nearest neighbor flux exchanges. Modifications are designed to make effective use of vectorization, streaming, and data management. Performance results using up to 16,384 graphics processing units of the Cray XK7 supercomputer Titan show more than 2.5× speedup over central processing unit-only performance on the same number of nodes (262,144 MPI ranks) for problem sizes of up to 6.9 billion grid points. We discuss performance-enhancement strategies and the overall potential of GPU-based computing for this class of problems.
Features, performance objectives and applications planned for the NASA National Aerodynamics Simulator (NAS) are outlined. NAS was conceived in the 1970s as a means to performing numerical aerodynamic simulations beyond the scope of wind tunnel testing for high-speed flight. Present supercomputers cannot deal with problems exhibiting strongly coupled viscous effects, which are being increasingly more accurately represented by approximations to the full Navier-Stokes equations. Located at the NASA-Ames Center, the NAS will by 1990 comprise a distributed computer network capable of a 4 Gflop computing rate and have a memory capacity of 1 billion 64-bit words. Remote access to the system through UNIX-based microcomputers will be available through land lines and satellite links. New supercomputers will be tested on the system without disturbing ongoing work. The core machine will be a Cray-2 with a 2 Gflop rate.
A class of implicit finite difference schemes of the Beam and Warming approximate factorization type will be addressed. The development and analysis of various aspects of this class of schemes will be given along with the motivations behind many of the choices. Various acceleration and efficiency modifications such as matrix reduction, diagonalization and flux split schemes will be presented. The methods are demonstrated in fully vectorized codes for a CRAY type architecture. The emphasis will be on the Euler equations in generalized coordinates.
We investigate a technique - segmental refinement (SR) - proposed by Brandt in the 1970s as a low memory multigrid method. The technique is attractive for modern computer architectures because it provides high data locality, minimizes network communication, is amenable to loop fusion, and is naturally highly parallel and asynchronous. The network communication minimization property was recognized by Brandt and Diskin in 1994; we continue this work by developing a segmental refinement method for a finite volume discretization of the 3D Laplacian on massively parallel computers. An understanding of the asymptotic complexities, required to maintain textbook multigrid efficiency, are explored experimentally with a simple SR method. A two-level memory model is developed to compare the asymptotic communication complexity of a proposed SR method with traditional parallel multigrid. Performance and scalability are evaluated with a Cray XC30 with up to 64K cores. We achieve modest improvement in scalability from traditional parallel multigrid with a simple SR implementation.
Computational prototyping of aircraft in nonlinear regimes will require the coupling of high-order disciplines on high performance computers. This effort investigates the coupling of the Reynolds-averaged Navier-Stokes equations with six degree-of-freedom nonlinear body dynamics and controls on distributed architectures. Towards this goal, solutions obtained on a wide-node IBM SP-2, an Ethernet-linked cluster of workstations, and a Cray C90 are compared in terms of accuracy and throughput.