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 199 records · Page 11

Supercomputing '91; Proceedings of the 4th Annual Conference on High Performance Computing, Albuquerque, NM, Nov. 18-22, 1991

Various papers on supercomputing are presented. The general topics addressed include: program analysis/data dependence, memory access, distributed memory code generation, numerical algorithms, supercomputer benchmarks, latency tolerance, parallel programming, applications, processor design, networks, performance tools, mapping and scheduling, characterization affecting performance, parallelism packaging, computing climate change, combinatorial algorithms, hardware and software performance issues, system issues. (No individual items are abstracted in this volume)

Source record↗

Porting Gravitational Wave Signal Extraction to Parallel Virtual Machine (PVM)

Laser Interferometer Space Antenna (LISA) is a planned NASA-ESA mission to be launched around 2012. The Gravitational Wave detection is fundamentally the determination of frequency, source parameters, and waveform amplitude derived in a specific order from the interferometric time-series of the rotating LISA spacecrafts. The LISA Science Team has developed a Mock LISA Data Challenge intended to promote the testing of complicated nested search algorithms to detect the 100-1 millihertz frequency signals at amplitudes of 10E-21. However, it has become clear that, sequential search of the parameters is very time consuming and ultra-sensitive; hence, a new strategy has been developed. Parallelization of existing sequential search algorithms of Gravitational Wave signal identification consists of decomposing sequential search loops, beginning with outermost loops and working inward. In this process, the main challenge is to detect interdependencies among loops and partitioning the loops so as to preserve concurrency. Existing parallel programs are based upon either shared memory or distributed memory paradigms. In PVM, master and node programs are used to execute parallelization and process spawning. The PVM can handle process management and process addressing schemes using a virtual machine configuration. The task scheduling and the messaging and signaling can be implemented efficiently for the LISA Gravitational Wave search process using a master and 6 nodes. This approach is accomplished using a server that is available at NASA Ames Research Center, and has been dedicated to the LISA Data Challenge Competition. Historically, gravitational wave and source identification parameters have taken around 7 days in this dedicated single thread Linux based server. Using PVM approach, the parameter extraction problem can be reduced to within a day. The low frequency computation and a proxy signal-to-noise ratio are calculated in separate nodes that are controlled by the master using message and vector of data passing. The message passing among nodes follows a pattern of synchronous and asynchronous send-and-receive protocols. The communication model and the message buffers are allocated dynamically to address rapid search of gravitational wave source information in the Mock LISA data sets.

Thirumalainambi, Rajkumar↗

Work stealing for GPU-accelerated parallel programs in a global address space framework

Task parallelism is an attractive approach to automatically load balance the computation in a parallel system and adapt to dynamism exhibited by parallel systems. Exploiting task parallelism through work stealing has been extensively studied in shared and distributed-memory contexts. In this paper, we study the design of a system that uses work stealing for dynamic load balancing of task-parallel programs executed on hybrid distributed-memory CPU-graphics processing unit (GPU) systems in a global-address space framework. We take into account the unique nature of the accelerator model employed by GPUs, the significant performance difference between GPU and CPU execution as a function of problem size, and the distinct CPU and GPU memory domains. We consider various alternatives in designing a distributed work stealing algorithm for CPU-GPU systems, while taking into account the impact of task distribution and data movement overheads. These strategies are evaluated using microbenchmarks that capture various execution configurations as well as the state-of-the-art CCSD(T) application module from the computational chemistry domain

97 MATHEMATICS AND COMPUTING↗

Scaling and Quality of Modularity Optimization Methods for Graph Clustering

Real-world graphs exhibit structures known as “communities” or “clusters” consisting of a group of vertices with relatively high connectivity between them, as compared to the rest of the vertices in the network. Graph clustering or community detection is a fundamental graph operation used to analyze real-world graphs occurring in the areas of computational biology, cybersecurity, electrical grids, etc. Similar to other graph algorithms, owing to irregular memory accesses and inherently sequential nature, current algorithms for community detection are challenging to parallelize. However, in order to analyze large networks, it is important to develop scalable parallel implementations of graph clustering that are capable of exploiting the architectural features of modern supercomputers. In response to the 2019 Streaming Graph Challenge, we present quality and performance analysis of our distributed- memory community detection using Vite, which is our distributed memory implementation of the popular Louvain method, on the ALCF Theta supercomputer. Clustering methods such as Louvain that rely on modularity maximization are known to suffer from the resolution limit problem, preventing identification of clusters of certain sizes. Hence, we also include quality analysis of our shared-memory implementation of the Fast-tracking Resistance method, in comparison with Louvain on the challenge datasets. Furthermore, we introduce an edge-balanced graph distribution for our distributed memory implementation, that significantly reduces communication, offering up to 80% improvement in the overall execution time. In addition to performance/quality analysis, we also include details on the power/energy consumption, and memory traffic of the distributed-memory clustering implementation using real-world graphs with over a billion edges.

Ghosh, Sayan↗

Parallel String Graph Construction and Transitive Reduction for De Novo Genome Assembly

One of the most computationally intensive tasks in computational biology is de novo genome assembly, the decoding of the sequence of an unknown genome from redundant and erroneous short sequences. A common assembly paradigm identifies overlapping sequences, simplifies their layout, and creates consensus. Despite many algorithms developed in the literature, the efficient assembly of large genomes is still an open problem. In this work, we introduce new distributed-memory parallel algorithms for overlap detection and layout simplification steps of de novo genome assembly, and implement them in the diBELLA 2D pipeline. Our distributed memory algorithms for both overlap detection and layout simplification are based on linear-algebra operations over semirings using 2D distributed sparse matrices. Our layout step consists of performing a transitive reduction from the overlap graph to a string graph. We provide a detailed communication analysis of the main stages of our new algorithms. diBELLA 2D achieves near linear scaling with over 80% parallel efficiency for the human genome, reducing the runtime for overlap detection by 1.2-1.3× for the human genome and 1.5-1.9× for C.elegans compared to the state-of-the-art. Our transitive reduction algorithm outperforms an existing distributed-memory implementation by 10.5-13.3× for the human genome and 18-29× for the C. elegans. Our work paves the way for efficient de novo assembly of large genomes using long reads in distributed memory.

59 BASIC BIOLOGICAL SCIENCES↗

Lifeline-based Global Load Balancing

On shared-memory systems, Cilk-style work-stealing has been used to effectively parallelize irregular task-graph based applications such as Unbalanced Tree Search (UTS). There are two main difficulties in extending this approach to distributed memory. In the shared memory approach, thieves (nodes without work) constantly attempt to asynchronously steal work from randomly chosen victims until they find work. In distributed memory, thieves cannot autonomously steal work from a victim without disrupting its execution. When work is sparse, this results in performance degradation. In essence, a direct extension of traditional work-stealing to distributed memory violates the work-first principle underlying work-stealing. Further, thieves spend useless CPU cycles attacking victims that have no work, resulting in system inefficiencies in multi-programmed contexts. Second, it is non-trivial to detect active distributed termination (detect that programs at all nodes are looking for work, hence there is no work). This problem is well-studied and requires careful design for good performance. Unfortunately, in most existing languages/frameworks, application developers are forced to implement their own distributed termination detection. In this paper, we develop a simple set of ideas that allow work-stealing to be efficiently extended to distributed memory. First, we introduce lifeline graphs: low-degree, low diameter, fully connected directed graphs. Such graphs can be constructed from k-dimensional hypercubes. When a node is unable to find work after w unsuccessful steals, it quiesces after informing the outgoing edges in its lifeline graph. Quiescent nodes do not disturb other nodes. A quiesced node is reactivated when work arrives from a lifeline and itself shares this work with those of its incoming lifelines that are activated. Termination occurs precisely when computation at all nodes has quiesced. In a language such as X10, such passive distributed termination can be detected automatically using the finish construct -- no application code is necessary.

99 GENERAL AND MISCELLANEOUS↗

Performance Modeling and Measurement of Parallelized Code for Distributed Shared Memory Multiprocessors

This paper presents a model to evaluate the performance and overhead of parallelizing sequential code using compiler directives for multiprocessing on distributed shared memory (DSM) systems. With increasing popularity of shared address space architectures, it is essential to understand their performance impact on programs that benefit from shared memory multiprocessing. We present a simple model to characterize the performance of programs that are parallelized using compiler directives for shared memory multiprocessing. We parallelized the sequential implementation of NAS benchmarks using native Fortran77 compiler directives for an Origin2000, which is a DSM system based on a cache-coherent Non Uniform Memory Access (ccNUMA) architecture. We report measurement based performance of these parallelized benchmarks from four perspectives: efficacy of parallelization process; scalability; parallelization overhead; and comparison with hand-parallelized and -optimized version of the same benchmarks. Our results indicate that sequential programs can conveniently be parallelized for DSM systems using compiler directives but realizing performance gains as predicted by the performance model depends primarily on minimizing architecture-specific data locality overhead.

Waheed, Abdul↗

Reducing Interprocessor Dependence in Recoverable Distributed Shared Memory

Checkpointing techniques in parallel systems use dependency tracking and/or message logging to ensure that a system rolls back to a consistent state. Traditional dependency tracking in distributed shared memory (DSM) systems is expensive because of high communication frequency. In this paper we show that, if designed correctly, a DSM system only needs to consider dependencies due to the transfer of blocks of data, resulting in reduced dependency tracking overhead and reduced potential for rollback propagation. We develop an ownership timestamp scheme to tolerate the loss of block state information and develop a passive server model of execution where interactions between processors are considered atomic. With our scheme, dependencies are significantly reduced compared to the traditional message-passing model.

Janssens, Bob↗

NAS Parallel Benchmark. Results 11-96: Performance Comparison of HPF and MPI Based NAS Parallel Benchmarks

High Performance Fortran (HPF), the high-level language for parallel Fortran programming, is based on Fortran 90. HALF was defined by an informal standards committee known as the High Performance Fortran Forum (HPFF) in 1993, and modeled on TMC's CM Fortran language. Several HPF features have since been incorporated into the draft ANSI/ISO Fortran 95, the next formal revision of the Fortran standard. HPF allows users to write a single parallel program that can execute on a serial machine, a shared-memory parallel machine, or a distributed-memory parallel machine. HPF eliminates the complex, error-prone task of explicitly specifying how, where, and when to pass messages between processors on distributed-memory machines, or when to synchronize processors on shared-memory machines. HPF is designed in a way that allows the programmer to code an application at a high level, and then selectively optimize portions of the code by dropping into message-passing or calling tuned library routines as 'extrinsics'. Compilers supporting High Performance Fortran features first appeared in late 1994 and early 1995 from Applied Parallel Research (APR) Digital Equipment Corporation, and The Portland Group (PGI). IBM introduced an HPF compiler for the IBM RS/6000 SP/2 in April of 1996. Over the past two years, these implementations have shown steady improvement in terms of both features and performance. The performance of various hardware/ programming model (HPF and MPI (message passing interface)) combinations will be compared, based on latest NAS (NASA Advanced Supercomputing) Parallel Benchmark (NPB) results, thus providing a cross-machine and cross-model comparison. Specifically, HPF based NPB results will be compared with MPI based NPB results to provide perspective on performance currently obtainable using HPF versus MPI or versus hand-tuned implementations such as those supplied by the hardware vendors. In addition we would also present NPB (Version 1.0) performance results for the following systems: DEC Alpha Server 8400 5/440, Fujitsu VPP Series (VX, VPP300, and VPP700), HP/Convex Exemplar SPP2000, IBM RS/6000 SP P2SC node (120 MHz) NEC SX-4/32, SGI/CRAY T3E, SGI Origin2000.

Saini, Subash↗

True Load Balancing for Matricized Tensor Times Khatri-Rao Product

MTTKRP is the bottleneck operation in algorithms used to compute the CP tensor decomposition. For sparse tensors, utilizing the compressed sparse fibers (CSF) storage format and the CSF-oriented MTTKRP algorithms is important for both memory and computational efficiency on distributed-memory architectures. Existing intelligent tensor partitioning models assume the computational cost of MTTKRP to be proportional to the total number of nonzeros in the tensor. However, this is not the case for the CSF-oriented MTTKRP on distributed-memory architectures. We outline two deficiencies of nonzero-based intelligent partitioning models when CSF-oriented MTTKRP operations are performed locally: failure to encode processors' computational loads and increase in total computation due to fiber fragmentation. We focus on existing fine-grain hypergraph model and propose a novel vertex weighting scheme that enables this model encode correct computational loads of processors. We also propose to augment the fine-grain model by fiber nets for reducing the increase in total computational load via minimizing fiber fragmentation. In this way, the proposed model encodes minimizing the load of the bottleneck processor. In conclusion, parallel experiments with real-world sparse tensors on up to 1024 processors prove the validity of the outlined deficiencies and demonstrate the merit of our proposed improvements in terms of parallel runtimes.

97 MATHEMATICS AND COMPUTING↗

Scaling Irregular Applications through Data Aggregation and Software Multithreading

Bioinformatics, data analytics, semantic databases, knowledge discovery are emerging high performance application areas that exploit dynamic, linked data structures such as graphs, unbalanced trees or unstructured grids. These data structures usually are very large, requiring significantly more memory than available on single shared memory systems. Additionally, these data structures are difficult to partition on distributed memory systems. They also present poor spatial and temporal locality, thus generating unpredictable memory and network accesses. The Partitioned Global Address Space (PGAS) programming model seems suitable for these applications, because it allows using a shared memory abstraction across distributed-memory clusters. However, current PGAS languages and libraries are built to target regular remote data accesses and block transfers. Furthermore, they usually rely on the Single Program Multiple Data (SPMD) parallel control model, which is not well suited to the fine grained, dynamic and unbalanced parallelism of irregular applications. In this paper we present {\bf GMT} (Global Memory and Threading library), a custom runtime library that enables efficient execution of irregular applications on commodity clusters. GMT integrates a PGAS data substrate with simple fork/join parallelism and provides automatic load balancing on a per node basis. It implements multi-level aggregation and lightweight multithreading to maximize memory and network bandwidth with fine-grained data accesses and tolerate long data access latencies. A key innovation in the GMT runtime is its thread specialization (workers, helpers and communication threads) that realize the overall functionality. We compare our approach with other PGAS models, such as UPC running using GASNet, and hand-optimized MPI code on a set of typical large-scale irregular applications, demonstrating speedups of an order of magnitude.

Aggregation↗

Aerodynamic Shape Optimization of Supersonic Aircraft Configurations via an Adjoint Formulation on Parallel Computers

This work describes the application of a control theory-based aerodynamic shape optimization method to the problem of supersonic aircraft design. The design process is greatly accelerated through the use of both control theory and a parallel implementation on distributed memory computers. Control theory is employed to derive the adjoint differential equations whose solution allows for the evaluation of design gradient information at a fraction of the computational cost required by previous design methods. The resulting problem is then implemented on parallel distributed memory architectures using a domain decomposition approach, an optimized communication schedule, and the MPI (Message Passing Interface) Standard for portability and efficiency. The final result achieves very rapid aerodynamic design based on higher order computational fluid dynamics methods (CFD). In our earlier studies, the serial implementation of this design method was shown to be effective for the optimization of airfoils, wings, wing-bodies, and complex aircraft configurations using both the potential equation and the Euler equations. In our most recent paper, the Euler method was extended to treat complete aircraft configurations via a new multiblock implementation. Furthermore, during the same conference, we also presented preliminary results demonstrating that this basic methodology could be ported to distributed memory parallel computing architectures. In this paper, our concern will be to demonstrate that the combined power of these new technologies can be used routinely in an industrial design environment by applying it to the case study of the design of typical supersonic transport configurations. A particular difficulty of this test case is posed by the propulsion/airframe integration.

Reuther, James↗

Aerodynamic Shape Optimization of Supersonic Aircraft Configurations via an Adjoint Formulation on Parallel Computers

This work describes the application of a control theory-based aerodynamic shape optimization method to the problem of supersonic aircraft design. The design process is greatly accelerated through the use of both control theory and a parallel implementation on distributed memory computers. Control theory is employed to derive the adjoint differential equations whose solution allows for the evaluation of design gradient information at a fraction of the computational cost required by previous design methods (13, 12, 44, 38). The resulting problem is then implemented on parallel distributed memory architectures using a domain decomposition approach, an optimized communication schedule, and the MPI (Message Passing Interface) Standard for portability and efficiency. The final result achieves very rapid aerodynamic design based on higher order computational fluid dynamics methods (CFD). In our earlier studies, the serial implementation of this design method (19, 20, 21, 23, 39, 25, 40, 41, 42, 43, 9) was shown to be effective for the optimization of airfoils, wings, wing-bodies, and complex aircraft configurations using both the potential equation and the Euler equations (39, 25). In our most recent paper, the Euler method was extended to treat complete aircraft configurations via a new multiblock implementation. Furthermore, during the same conference, we also presented preliminary results demonstrating that the basic methodology could be ported to distributed memory parallel computing architectures [241. In this paper, our concem will be to demonstrate that the combined power of these new technologies can be used routinely in an industrial design environment by applying it to the case study of the design of typical supersonic transport configurations. A particular difficulty of this test case is posed by the propulsion/airframe integration.

Reuther, James↗

Mapping robust parallel multigrid algorithms to scalable memory architectures

The convergence rate of standard multigrid algorithms degenerates on problems with stretched grids or anisotropic operators. The usual cure for this is the use of line or plane relaxation. However, multigrid algorithms based on line and plane relaxation have limited and awkward parallelism and are quite difficult to map effectively to highly parallel architectures. Newer multigrid algorithms that overcome anisotropy through the use of multiple coarse grids rather than line relaxation are better suited to massively parallel architectures because they require only simple point-relaxation smoothers. The parallel implementation of a V-cycle multiple semi-coarsened grid (MSG) algorithm or distributed-memory architectures such as the Intel iPSC/860 and Paragon computers is addressed. The MSG algorithms provide two levels of parallelism: parallelism within the relaxation or interpolation on each grid and across the grids on each multigrid level. Both levels of parallelism must be exploited to map these algorithms effectively to parallel architectures. A mapping of an MSG algorithm to distributed-memory architectures that demonstrate how both levels of parallelism can be exploited is described. The results is a robust and effective multigrid algorithm for distributed-memory machines.

Overman, Andrea↗

Mapping robust parallel multigrid algorithms to scalable memory architectures

The convergence rate of standard multigrid algorithms degenerates on problems with stretched grids or anisotropic operators. The usual cure for this is the use of line or plane relaxation. However, multigrid algorithms based on line and plane relaxation have limited and awkward parallelism and are quite difficult to map effectively to highly parallel architectures. Newer multigrid algorithms that overcome anisotropy through the use of multiple coarse grids rather than relaxation are better suited to massively parallel architectures because they require only simple point-relaxation smoothers. In this paper, we look at the parallel implementation of a V-cycle multiple semicoarsened grid (MSG) algorithm on distributed-memory architectures such as the Intel iPSC/860 and Paragon computers. The MSG algorithms provide two levels of parallelism: parallelism within the relaxation or interpolation on each grid and across the grids on each multigrid level. Both levels of parallelism must be exploited to map these algorithms effectively to parallel architectures. This paper describes a mapping of an MSG algorithm to distributed-memory architectures that demonstrates how both levels of parallelism can be exploited. The result is a robust and effective multigrid algorithm for distributed-memory machines.

Overman, Andrea↗

An Asynchronous Task-based Fan-Both Sparse Cholesky Solver

Systems of linear equations arise at the heart of many scientific and engineering applications. Many of these linear systems are sparse; i.e., most of the elements in the coefficient matrix are zero. Direct methods based on matrix factorizations are sometimes needed to ensure accurate solutions. For example, accurate solution of sparse linear systems is needed in shift-invert Lanczos to compute interior eigenvalues. The performance and resource usage of sparse matrix factorizations are critical to time-to-solution and maximum problem size solvable on a given platform. In many applications, the coefficient matrices are symmetric, and exploiting symmetry will reduce both the amount of work and storage cost required for factorization. When the factorization is performed on large-scale distributed memory platforms, communication cost is critical to the performance of the algorithm. At the same time, network topologies have become increasingly complex, so that modern platforms exhibit a high level of performance variability. This makes scheduling of computations an intricate and performance-critical task. In this paper, we investigate the use of an asynchronous task paradigm, one-sided communication and dynamic scheduling in implementing sparse Cholesky factorization (symPACK) on large-scale distributed memory platforms. Our solver symPACK relies on efficient and flexible communication primitives provided by the UPC++ library. Performance evaluation shows good scalability and that symPACK outperforms state-of-the-art parallel distributed memory factorization packages, validating our approach on practical cases.

Jacquelin, Mathias↗

Cross-scale efficient tensor contractions for coupled cluster computations through multiple programming model backends

Coupled-cluster methods provide highly accurate models of molecular structure through explicit numerical calculation of tensors representing the correlation between electrons. These calculations are dominated by a sequence of tensor contractions, motivating the development of numerical libraries for such operations. While based on matrix–matrix multiplication, these libraries are specialized to exploit symmetries in the molecular structure and in electronic interactions, and thus reduce the size of the tensor representation and the complexity of contractions. The resulting algorithms are irregular and their parallelization has been previously achieved via the use of dynamic scheduling or specialized data decompositions. We introduce our efforts to extend the Libtensor framework to work in the distributed memory environment in a scalable and energy-efficient manner. We achieve up to 240× speedup compared with the optimized shared memory implementation of Libtensor. We attain scalability to hundreds of thousands of compute cores on three distributed-memory architectures (Cray XC30 and XC40, and IBM Blue Gene/Q), and on a heterogeneous GPU-CPU system (Cray XK7). As the bottlenecks shift from being compute-bound DGEMM's to communication-bound collectives as the size of the molecular system scales, we adopt two radically different parallelization approaches for handling load-imbalance, tasking and bulk synchronous models. Nevertheless, we preserve a unified interface to both programming models to maintain the productivity of computational quantum chemists.

97 MATHEMATICS AND COMPUTING↗

Weather prediction using a genetic memory

Kanaerva's sparse distributed memory (SDM) is an associative memory model based on the mathematical properties of high dimensional binary address spaces. Holland's genetic algorithms are a search technique for high dimensional spaces inspired by evolutional processes of DNA. Genetic Memory is a hybrid of the above two systems, in which the memory uses a genetic algorithm to dynamically reconfigure its physical storage locations to reflect correlations between the stored addresses and data. This architecture is designed to maximize the ability of the system to scale-up to handle real world problems.

Rogers, David↗