Search NASA⌕ Search

SEARCH · Search NASA

Results for “distributed memory”

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.

At least 523 records · Page 29

Application Characterization at Scale: Lessons learned from developing a distributed Open Community Runtime system for High Performance Computing

Since 2012, the U.S. Department of Energy’s X-Stack program has been developing solutions including runtime systems, programming models, languages, compilers, and tools for the Exascale system software to address crucial performance and power requirements. Fine grain programming models and runtime systems show a great potential to efficiently utilize the underlying hardware. Thus, they are essential to many X-Stack efforts. An abundant amount of small tasks can better utilize the vast parallelism available on current and future machines. Moreover, finer tasks can recover faster and adapt better, due to a decrease in state and control. Nevertheless, current applications have been written to exploit old paradigms (such as Communicating Sequential Processor and Bulk Synchronous Parallel processing). To fully utilize the advantages of these new systems, applications need to be adapted to these new paradigms. As part of the applications’ porting process, in-depth characterization studies, focused on both application characteristics and runtime features, need to take place to fully understand the application performance bottlenecks and how to resolve them. This paper presents a characterization study for a novel high performance runtime system, called the Open Community Runtime, using key HPC kernels as its vehicle. This study has the following contributions: one of the first high performance, fine grain, distributed memory runtime system implementing the OCR standard (version 0.99a); and a characterization study of key HPC kernels in terms of runtime primitives running on both intra and inter node environments. Running on a general purpose cluster, we have found up to 1635x relative speed-up for a parallel tiled Cholesky Kernels on 128 nodes with 16 cores each and a 1864x relative speed-up for a parallel tiled Smith-Waterman kernel on 128 nodes with 30 cores.

Application characterization↗

What does fault tolerant Deep Learning need from MPI?

Deep Learning (DL) algorithms have become the {\em de facto} Machine Learning (ML) algorithm for large scale data analysis. DL algorithms are computationally expensive -- even distributed DL implementations which use MPI require days of training (model learning) time on commonly studied datasets. Long running DL applications become susceptible to faults -- requiring development of a fault tolerant system infrastructure, in addition to fault tolerant DL algorithms. This raises an important question: {\em What is needed from MPI for designing fault tolerant DL implementations?} In this paper, we address this problem for permanent faults. We motivate the need for a fault tolerant MPI specification by an in-depth consideration of recent innovations in DL algorithms and their properties, which drive the need for specific fault tolerance features. We present an in-depth discussion on the suitability of different parallelism types (model, data and hybrid); a need (or lack thereof) for check-pointing of any critical data structures; and most importantly, consideration for several fault tolerance proposals (user-level fault mitigation (ULFM), Reinit) in MPI and their applicability to fault tolerant DL implementations. We leverage a distributed memory implementation of Caffe, currently available under the Machine Learning Toolkit for Extreme Scale (MaTEx). We implement our approaches by extending MaTEx-Caffe for using ULFM-based implementation. Our evaluation using the ImageNet dataset and AlexNet neural network topology demonstrates the effectiveness of the proposed fault tolerant DL implementation using OpenMPI based ULFM.

Amatya, Vinay C.↗

Scalable Cloning on Large-Scale GPU Platforms with Application to Time-Stepped Simulations on Grids

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.

97 MATHEMATICS AND COMPUTING↗

Integrated Model, Batch and Domain Parallelism in Training Neural Networks

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.

97 MATHEMATICS AND COMPUTING↗

A communication-avoiding 3D sparse triangular solver

We present a novel distributed memory algorithm to improve the strong scalability of the solution of a sparse triangular system. This operation appears in the solve phase of direct methods for solving general sparse linear systems, Ax = b. Our 3D sparse triangular solver employs several techniques, including a 3D MPI process grid, elimination tree parallelism, and data replication, all of which reduce the per-process communication when combined. We present analytical models to understand the communication cost of our algorithm and show that our 3D sparse triangular solver can reduce the per-process communication volume asymptotically by a factor of O(n1/4) and O(n1/6) for problems arising from the finite element discretizations of 2D "planar" and 3D "non-planar" PDEs, respectively. We implement our algorithm for use in SuperLU_DIST3D, using a hybrid MPI+OpenMP programming model. Our 3D triangular solve algorithm, when run on 12k cores of Cray XC30, outperforms the current state-of-the-art 2D algorithm by 7.2x for planar and 2.7x for the non-planar sparse matrices, respectively.

Sao, Piyush↗

Performance Models for Data Transfers: A Case Study with Molecular Chemistry Kernels

In distributed memory systems, it is paramount to develop strategies to overlap the data transfers between memory nodes with the computations in order to exploit the full potential of these systems. In this paper, we consider the problem of determining the order of data transfers between two memory nodes for a set of independent tasks with the objective of minimizing the makespan. We prove that, with limited memory capacity, the problem of obtaining the optimal order of data transfers is NP-complete. We propose several heuristics to determine this order and discuss the conditions that might be favorable to different heuristics. We analyze our heuristics on traces obtained by running 2 molecular chemistry kernels, namely, Hartree–Fock (HF) and Coupled Cluster Single Double (CCSD) on 10 nodes of an HPC system. Our results show that some of our heuristics achieve significant overlap for moderate memory capacities and resulting in makespans that are very close to the lower bound.

Kumar, Suraj↗

Simulating the Impact of Dynamic Rerouting on Metropolitan-scale Traffic Systems

The rapid introduction of mobile navigation aides that use real-time road network information to suggest alternate routes to drivers is making it more difficult for researchers and government transportation agencies to understand and predict the dynamics of congested transportation systems. Computer simulation is a key capability for these organizations to analyze hypothetical scenarios; however, the complexity of transportation systems makes it challenging for them to simulate very large geographical regions, such as multi-city metropolitan areas. In this article, we describe enhancements to the Mobiliti parallel traffic simulator to model dynamic rerouting behavior with the addition of vehicle controller actors and vehicle-to-controller reroute requests. The simulator is designed to support distributed-memory parallel execution using discrete event simulation and be scalable on high-performance computing platforms. We demonstrate the potential of the simulator by analyzing the impact of varying the population penetration rate of dynamic rerouting on the San Francisco Bay Area road network. Using high-performance parallel computing, we can simulate a day in the San Francisco Bay Area with 19 million vehicle trips with 50 percent dynamic rerouting penetration over a road network with 0.5 million nodes and 1 million links in less than three minutes. We present a sensitivity study on the dynamic rerouting parameters, discuss the simulator’s parallel scalability, and analyze system-level impacts of changing the dynamic rerouting penetration. Furthermore, we examine the varying effects on different functional classes and geographical regions and present a validation of the simulation results compared to real-world data.

97 MATHEMATICS AND COMPUTING↗

Performance Analysis and Optimal Node-aware Communication for Enlarged Conjugate Gradient Methods

Krylov methods are a key way of solving large sparse linear systems of equations but suffer from poor strong scalability on distributed memory machines. Furthermore, this is due to high synchronization costs from large numbers of collective communication calls alongside a low computational workload. Enlarged Krylov methods address this issue by decreasing the total iterations to convergence, an artifact of splitting the initial residual and resulting in operations on block vectors. In this article, we present a performance study of an enlarged Krylov method, Enlarged Conjugate Gradients (ECG), noting the impact of block vectors on parallel performance at scale. Most notably, we observe the increased overhead of point-to-point communication as a result of denser messages in the sparse matrix-block vector multiplication kernel. Additionally, we present models to analyze expected performance of ECG, as well as motivate design decisions. Most importantly, we introduce a new point-to-point communication approach based on node-aware communication techniques that increases efficiency of the method at scale.

97 MATHEMATICS AND COMPUTING↗

Communication Lower Bounds and Optimal Algorithms for Symmetric Matrix Computations

In this article, we focus on the communication costs of three symmetric matrix computations: (i) multiplying a matrix with its transpose, known as a symmetric rank-k update (SYRK) (ii) adding the result of the multiplication of a matrix with the transpose of another matrix and the transpose of that result, known as a symmetric rank-2k update (SYR2K) (iii) performing matrix multiplication with a symmetric input matrix (SYMM). All three computations appear in the Level 3 Basic Linear Algebra Subroutines (BLAS) and have wide use in applications involving symmetric matrices. We establish communication lower bounds for these kernels using sequential and distributed-memory parallel computational models, and we show that our bounds are tight by presenting communication-optimal algorithms for each setting. Our lower bound proofs rely on applying a geometric inequality for symmetric computations and analytically solving constrained nonlinear optimization problems. As a result, the symmetric matrix and its corresponding computations are accessed and performed according to a triangular block partitioning scheme in the optimal algorithms.

Al Daas, Hussam [Rutherford Appleton Laboratory, D↗

Integrated Network Decompositions and Dynamic Programming for Graph Optimization (INDDGO)

The INDDGO software package offers a set of tools for finding exact solutions to graph optimization problems via tree decompositions and dynamic programming algorithms. Currently the framework offers serial and parallel (distributed memory) algorithms for finding tree decompositions and solving the maximum weighted independent set problem. The parallel dynamic programming algorithm is implemented on top of the MADNESS task-based runtime.

Source record↗

OpenSHMEM over MPI-3

OpenSHMEM is a community effort to unifyt and standardize the SHMEM programming model. MPI (Message Passing Interface) is a well-known community standard for parallel programming using distributed memory. The most recen t release of MPI, version 3.0, was designed in part to support programming models like SHMEM.OSHMPI is an implementation of the OpenSHMEM standard using MPI-3 for the Linux operating system. It is the first implementation of SHMEM over MPI one-sided communication and has the potential to be widely adopted due to the portability and widely availability of Linux and MPI-3. OSHMPI has been tested on a variety of systems and implementations of MPI-3, includingInfiniBand clusters using MVAPICH2 and SGI shared-memory supercomputers using MPICH. Current support is limited to Linux but may be extended to Apple OSX if there is sufficient interest. The code is opensource via https://github.com/jeffhammond/oshmpi

Sayan Ghosh, JeffHammond↗

Parallel Element Agglomeration Algebraic Multigrid and Upscaling Library

ParFELAG is a parallel distributed memory C++ library for numerical upscaling of finite element discretizations. It provides optimal complesity algorithms ro build multilevel hierarchies and solvers that can be used for solving a wide class of partial differential equations (elliptic, hyperbolic, saddle point problems) on general unstructured mesh (under the assumption that the topology of the agglomerated entities is correct). Additionally, a novel multilevel solver for saddle point problems with divergence constraint is implemented.

Source record↗

PeleC [SWR-17-35]

PeleC is an adaptive-mesh compressible hydrodynamics code for reacting flows. It solves the compressible Navier-Stokes with multispecies transport in a block structured framework. The resulting algorithm is well suited for flows with localized resolution requirements and robust to discontinuities. User controllable refinement crieteria has the potential to result in extremely small numerical dissipation and dispersion, making this code appropriate for both research and applied usage. The code is built on the AMReX library which facilitates hierarchical parallelism and manages distributed memory parallism. PeleC algorithms are implemented to express shared memory parallelism.

Source record↗

MapReduce MPI library (MR-MPI) v. 1.0

The MapReduce MPI library is a software tool for performing MapReduce operations on a distributed memory parallel computer via message passing (MPI). These are data-processing or computational operations which achieve parallelism by breaking a large task into two stages, a map and a reduce . Each of these are formulated as simple on-processor functions which the user can easily write. The library assigns independent tasks to processors and performs the parallel data movement and transformations behind the scenes.

PLIMPTON, STEVEN↗

Parallel Element Agglomeration Multilevel Monte Carlo Library (ParELAGMC)

ParELAGMC is a parallel distributed memory C++ library for multilevel Monte Carlo (MLMC) simulations with algebraically constructed coarse spaces for estimating statistical moments of a quantity of interest involving solutions of partial differential equations with uncertain coefficients. A primary focus of the library is the scalable generation of Gaussian random fields using a novel stochastic PDE (SPDE) sampling technique for large-scale simulations. A key feature in the SPDE sampling technique is a novel non-matching mesh embedding technique to mitigate artificial boundary artifacts that are inherent in the method. Example applications to uncertainty quantification of subsurface flow simulations in the mixed finite element setting are provided as well.

Osborn, SarahV.↗

ROVER

ROVER Is an application for distributed-memory volume rendering and simulated radiography focused on many-core architectures (e.g., GPU, CPU, and MIC). Volume rendering, also known as volume visualization, is a commonly used scientific visualization algorithm that enables users to visualize an entire scalar field at once by mapping scalar values to colors and transparency. Simulated radiography Is a generalization of the volume rendering algorithm from four color channels (I.e., red, green, blue, and transparency) to an arbitrary number of channels representing material opacities in discrete energy groups, measured In electron volts.

Larsen, Matthew↗

Parallel Markov Random Fields for Image Segmentation (PMRF-IS) v1.0

This software performs multi-label segmentation of 2D/3D images. It is highly accurate, results rely on use of an Markov random field (MRF) formulation. It runs in both shared and distributed memory parallel modes. MRF algorithms are powerful tools in image analysis to explore contextual information of data.

Heinemann, Colleen↗

Berkeley Unified Parallel C (BUPC) Compiler

This program is a portable, open-source, compiler for the UPC language. It is based on the Open64 framework and has extensive support for optimizations. This compiler operates by translating UPC into ANSI/ISO C for compilation by a native compiler and linking with the UPC Runtime Library. This design eases portability to both shared- and distributed-memory parallel architectures. For proper operation it requires the "Berkeley Unified Parallel C (UPC) Runtime Library" and its dependencies.

Iancu, Costin↗