Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “distributed memory parallel”

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 451 records · Page 25

Hierarchical resilience with lightweight threads.

This paper proposes methodology for providing robustness and resilience for a highly threaded distributed- and shared-memory environment based on well-defined inputs and outputs to lightweight tasks. These inputs and outputs form a failure 'barrier', allowing tasks to be restarted or duplicated as necessary. These barriers must be expanded based on task behavior, such as communication between tasks, but do not prohibit any given behavior. One of the trends in high-performance computing codes seems to be a trend toward self-contained functions that mimic functional programming. Software designers are trending toward a model of software design where their core functions are specified in side-effect free or low-side-effect ways, wherein the inputs and outputs of the functions are well-defined. This provides the ability to copy the inputs to wherever they need to be - whether that's the other side of the PCI bus or the other side of the network - do work on that input using local memory, and then copy the outputs back (as needed). This design pattern is popular among new distributed threading environment designs. Such designs include the Barcelona STARS system, distributed OpenMP systems, the Habanero-C and Habanero-Java systems from Vivek Sarkar at Rice University, the HPX/ParalleX model from LSU, as well as our own Scalable Parallel Runtime effort (SPR) and the Trilinos stateless kernels. This design pattern is also shared by CUDA and several OpenMP extensions for GPU-type accelerators (e.g. the PGI OpenMP extensions).

43 PARTICLE ACCELERATORS↗

SPARC: Demonstrate burst-buffer-based checkpoint/restart on ATS-1.

Recent high-performance computing (HPC) platforms such as the Trinity Advanced Technology System (ATS-1) feature burst buffer resources that can have a dramatic impact on an application’s I/O performance. While these non-volatile memory (NVM) resources provide a new tier in the storage hierarchy, developers must find the right way to incorporate the technology into their applications in order to reap the benefits. Similar to other laboratories, Sandia is actively investigating ways in which these resources can be incorporated into our existing libraries and workflows without burdening our application developers with excessive, platform-specific details. This FY18Q1 milestone summaries our progress in adapting the Sandia Parallel Aerodynamics and Reentry Code (SPARC) in Sandia’s ATDM program to leverage Trinity’s burst buffers for checkpoint/restart operations. We investigated four different approaches with varying tradeoffs in this work: (1) simply updating job script to use stage-in/stage out burst buffer directives, (2) modifying SPARC to use LANL’s hierarchical I/O (HIO) library to store/retrieve checkpoints, (3) updating Sandia’s IOSS library to incorporate the burst buffer in all meshing I/O operations, and (4) modifying SPARC to use our Kelpie distributed memory library to store/retrieve checkpoints. Team members were successful in generating initial implementation for all four approaches, but were unable to obtain performance numbers in time for this report (reasons: initial problem sizes were not large enough to stress I/O, and SPARC refactor will require changes to our code). When we presented our work to the SPARC team, they expressed the most interest in the second and third approaches. The HIO work was favored because it is lightweight, unobtrusive, and should be portable to ATS-2. The IOSS work is seen as a long-term solution, and is favored because all I/O work (including checkpoints) can be deferred to a single library.

97 MATHEMATICS AND COMPUTING↗

Modeling the asynchronous Jacobi method without communication delays

Asynchronous iterative methods for solving linear systems are gaining renewed interest due to the high cost of synchronization points in massively parallel codes. Historically, theory on asynchronous iterative methods has focused on asymptotic behavior, while the transient behavior remains poorly understood. In this paper, we study a model of the asynchronous Jacobi method without communication delays, which we call simplified asynchronous Jacobi. Simplified asynchronous Jacobi can be used to model asynchronous Jacobi implemented in shared memory or distributed memory with fast communication networks. Our analysis uses the idea of a propagation matrix, which is similar in concept to an iteration matrix. We show that simplified asynchronous Jacobi can continue to reduce the residual when some processes are slower than other processes. We also show that simplified asynchronous Jacobi can converge when synchronous Jacobi does not. We verify our analysis of simplified asynchronous Jacobi using results from asynchronous Jacobi implemented in shared and distributed memory.

97 MATHEMATICS AND COMPUTING↗

Massively Parallel Quantum Chemistry: A high-performance research platform for electronic structure

The Massively Parallel Quantum Chemistry (MPQC) program is a 30-year-old project that enables facile development of electronic structure methods for molecules for efficient deployment to massively parallel computing architectures. Here, we describe the historical evolution of MPQC’s design into its latest (fourth) version, the capabilities and modular architecture of today’s MPQC, and how MPQC facilitates rapid composition of new methods as well as its state-of-the-art performance on a variety of commodity and high-end distributed-memory computer platforms.

Peng, Chong↗

Supporting 64-bit global indices in Epetra and other Trilinos packages :

The Trilinos Project is an effort to facilitate the design, development, integration and ongoing support of mathematical software libraries within an object-oriented framework. It is intended for large-scale, complex multiphysics engineering and scientific applications [2, 4, 3]. Epetra is one of its basic packages. It provides serial and parallel linear algebra capabilities. Before Trilinos version 11.0, released in 2012, Epetra used the C++ int data-type for storing global and local indices for degrees of freedom (DOFs). Since int is typically 32-bit, this limited the largest problem size to be smaller than approximately two billion DOFs. This was true even if a distributed memory machine could handle larger problems. We have added optional support for C++ long long data-type, which is at least 64-bit wide, for global indices. To save memory, maintain the speed of memory-bound operations, and reduce further changes to the code, the local indices are still 32-bit. We document the changes required to achieve this feature and how the new functionality can be used. We also report on the lessons learned in modifying a mature and popular package from various perspectives design goals, backward compatibility, engineering decisions, C++ language features, effects on existing users and other packages, and build integration.

Jhurani, Chetan↗

Optimizing Distributed Training on Frontier for Large Language Models

Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Loss scaling studies have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of 38.38%, 36.14%, and 31.96%, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved 100% weak scaling efficiency on 1024 and 3072 Mi250X GPUs, respectively. We also achieved strong scaling efficiencies of 89% and 87% for these two models. We trained these models only tens of iterations instead of training till completion.

Yin, Junqi↗

Asymptotic Behavior of Memristive Circuits

The interest in memristors has risen due to their possible application both as memory units and as computational devices in combination with CMOS. This is in part due to their nonlinear dynamics, and a strong dependence on the circuit topology. We provide evidence that also purely memristive circuits can be employed for computational purposes. In the present paper we show that a polynomial Lyapunov function in the memory parameters exists for the case of DC controlled memristors. Such a Lyapunov function can be asymptotically approximated with binary variables, and mapped to quadratic combinatorial optimization problems. This also shows a direct parallel between memristive circuits and the Hopfield-Little model. In the case of Erdos-Renyi random circuits, we show numerically that the distribution of the matrix elements of the projectors can be roughly approximated with a Gaussian distribution, and that it scales with the inverse square root of the number of elements. This provides an approximated but direct connection with the physics of disordered system and, in particular, of mean field spin glasses. Using this and the fact that the interaction is controlled by a projector operator on the loop space of the circuit. We estimate the number of stationary points of the approximate Lyapunov function and provide a scaling formula as an upper bound in terms of the circuit topology only.

97 MATHEMATICS AND COMPUTING↗

Tuning collective communication for Partitioned Global Address Space programming models

Partitioned Global Address Space (PGAS) languages offer programmers the convenience of a shared memory programming style combined with locality control necessary to run on large-scale distributed memory systems. Even within a PGAS language programmers often need to perform global communication operations such as broadcasts or reductions, which are best performed as collective operations in which a group of threads work together to perform the operation. In this study we consider the problem of implementing collective communication within PGAS languages and explore some of the design trade-offs in both the interface and implementation. In particular, PGAS collectives have semantic issues that are different than in send–receive style message passing programs, and different implementation approaches that take advantage of the one-sided communication style in these languages. We present an implementation framework for PGAS collectives as part of the GASNet communication layer, which supports shared memory, distributed memory and hybrids. The framework supports a broad set of algorithms for each collective, over which the implementation may be automatically tuned. In conclusion, we demonstrate the benefit of optimized GASNet collectives using application benchmarks written in UPC, and demonstrate that the GASNet collectives can deliver scalable performance on a variety of state-of-the-art parallel machines including a Cray XT4, an IBM BlueGene/P, and a Sun Constellation system with InfiniBand interconnect.

97 MATHEMATICS AND COMPUTING↗

Compressed basis GMRES on high-performance graphics processing units

Krylov methods provide a fast and highly parallel numerical tool for the iterative solution of many large-scale sparse linear systems. To a large extent, the performance of practical realizations of these methods is constrained by the communication bandwidth in current computer architectures, motivating the investigation of sophisticated techniques to avoid, reduce, and/or hide the message-passing costs (in distributed platforms) and the memory accesses (in all architectures). This article leverages Ginkgo’s memory accessor in order to integrate a communication-reduction strategy into the (Krylov) GMRES solver that decouples the storage format (i.e., the data representation in memory) of the orthogonal basis from the arithmetic precision that is employed during the operations with that basis. Given that the execution time of the GMRES solver is largely determined by the memory accesses, the cost of the datatype transforms can be mostly hidden, resulting in the acceleration of the iterative step via a decrease in the volume of bits being retrieved from memory. Together with the special properties of the orthonormal basis (whose elements are all bounded by 1), this paves the road toward the aggressive customization of the storage format, which includes some floating-point as well as fixed-point formats with mild impact on the convergence of the iterative process. We develop a high-performance implementation of the “compressed basis GMRES” solver in the Ginkgo sparse linear algebra library using a large set of test problems from the SuiteSparse Matrix Collection. We demonstrate robustness and performance advantages on a modern NVIDIA V100 graphics processing unit (GPU) of up to 50% over the standard GMRES solver that stores all data in IEEE double-precision.

97 MATHEMATICS AND COMPUTING↗

Towards Superior Software Portability with SHAD and HPX C++ Libraries

As hardware architectures and software stacks complexity grows, development productivity, performance and software portability, quickly evolve from desirable features to actual needs. SHAD, the Scalable High-performance Algorithms and Data-structures C++ library is designed to mitigate these issues: it provides general purpose building blocks as well as high-level custom utilities, and offers a shared-memory programming abstraction which facilitates the programming of complex systems, scaling up to High Performance Computing clusters. SHAD’s portability is achieved through an abstract runtime interface, which decouples the upper layers of the library and hides the low level details of the underlying architecture. This layer enables SHAD to interface with different runtime/threading systems, e.g. Intel TBB and Global Memory and Threading (GMT). However, current backends targeting distributed systems, rely on a centralized controller which may possibly limit scalability up to hundreds of nodes and creates a network hot spot due to all to one communication for synchronization, and possibly resulting in degraded performance at high process counts. In this research, we explore HPX, the C++ standard library for parallelism and concurrency, as an additional backend in support of the SHAD library, and present the methodologies in support of local and remote task executions in SHAD with respect to HPX. Finally, we evaluate the proposed system by comparing against existing backends of SHAD and analyzing their performance on C++ Standard Template Library algorithms.

Wu, Nanmiao↗

Structural Simluation Toolkit (SST) V 2.0

The SST provides a parallel framework to perform system simulation of computer architectures to determine their performance and power consumption. Additionally, the SST contains basic models of a computer processor, and interconnect and can connect to an external memory simulator (DRAMSim II). The SST framework provides a simple interface by which other computer simulation models can be combined under a common parallel discrete event-based simulation environment. This allows design exploration of future architectures, analysis of how current computer programs will function on future architectures. The SST provides a parallel discrete event simulation framework, including partitioning and object distribution over MPI. It also provides a mechanism by which components can report their power consumption for analysis.

RODRIGUES, ARUN↗

Structural Simluation Toolkit (SST) v.11.0

The SST provides a parallel framework to perform system simulation of computer architectures to determine their performance and power consumption. Additionally, the SST contains basic models of a computer processor, and interconnect and can connect to an external memory simulator (DRAMSim II). The SST framework provides a simple interface by which other computer simulation models can be combined under a common parallel discrete event-based simulation environment. This allows design exploration of future architectures, analysis of how current computer programs will function on future architectures. The SST provides a parallel discrete event simulation framework, including partitioning and object distribution over MPI. It also provides a mechanism by which components can report their power consumption for analysis.

Rodrigues, ArunF.↗

Scalable Performance Measurement and Analysis

Concurrency levels in large-scale, distributed-memory supercomputers are rising exponentially. Modern machines may contain 100,000 or more microprocessor cores, and the largest of these, IBM's Blue Gene/L, contains over 200,000 cores. Future systems are expected to support millions of concurrent tasks. In this dissertation, we focus on efficient techniques for measuring and analyzing the performance of applications running on very large parallel machines. Tuning the performance of large-scale applications can be a subtle and time-consuming task because application developers must measure and interpret data from many independent processes. While the volume of the raw data scales linearly with the number of tasks in the running system, the number of tasks is growing exponentially, and data for even small systems quickly becomes unmanageable. Transporting performance data from so many processes over a network can perturb application performance and make measurements inaccurate, and storing such data would require a prohibitive amount of space. Moreover, even if it were stored, analyzing the data would be extremely time-consuming. In this dissertation, we present novel methods for reducing performance data volume. The first draws on multi-scale wavelet techniques from signal processing to compress systemwide, time-varying load-balance data. The second uses statistical sampling to select a small subset of running processes to generate low-volume traces. A third approach combines sampling and wavelet compression to stratify performance data adaptively at run-time and to reduce further the cost of sampled tracing. We have integrated these approaches into Libra, a toolset for scalable load-balance analysis. We present Libra and show how it can be used to analyze data from large scientific applications scalably.

99 GENERAL AND MISCELLANEOUS↗

Hybrid and Parallel Domain-Decomposition Methods Development to Enable Monte Carlo for Reactor Analyses

This paper describes code and methods development at the Oak Ridge National Laboratory focused on enabling high-fidelity, large-scale reactor analyses with Monte Carlo (MC). Current state-of-the-art tools and methods used to perform real commercial reactor analyses have several undesirable features, the most significant of which is the non-rigorous spatial decomposition scheme. Monte Carlo methods, which allow detailed and accurate modeling of the full geometry and are considered the gold standard for radiation transport solutions, are playing an ever-increasing role in correcting and/or verifying the deterministic, multi-level spatial decomposition methodology in current practice. However, the prohibitive computational requirements associated with obtaining fully converged, system-wide solutions restrict the role of MC to benchmarking deterministic results at a limited number of state-points for a limited number of relevant quantities. The goal of this research is to change this paradigm by enabling direct use of MC for full-core reactor analyses. The most significant of the many technical challenges that must be overcome are the slow, non-uniform convergence of system-wide MC estimates and the memory requirements associated with detailed solutions throughout a reactor (problems involving hundreds of millions of different material and tally regions due to fuel irradiation, temperature distributions, and the needs associated with multi-physics code coupling). To address these challenges, our research has focused on the development and implementation of (1) a novel hybrid deterministic/MC method for determining high-precision fluxes throughout the problem space in k-eigenvalue problems and (2) an efficient MC domain-decomposition (DD) algorithm that partitions the problem phase space onto multiple processors for massively parallel systems, with statistical uncertainty estimation. The hybrid method development is based on an extension of the FW-CADIS method, which attempts to achieve uniform statistical uncertainty throughout a designated problem space. The MC DD development is being implemented in conjunction with the Denovo deterministic radiation transport package to have direct access to the 3-D, massively parallel discrete-ordinates solver (to support the hybrid method) and the associated parallel routines and structure. This paper describes the hybrid method, its implementation, and initial testing results for a realistic 2-D quarter core pressurized-water reactor model and also describes the MC DD algorithm and its implementation.

21 SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLANTS↗

Hybrid and Parallel Domain-Decomposition Methods Development to Enable Monte Carlo for Reactor Analyses

This paper describes code and methods development at the Oak Ridge National Laboratory focused on enabling high-fidelity, large-scale reactor analyses with Monte Carlo (MC). Current state-of-the-art tools and methods used to perform ''real'' commercial reactor analyses have several undesirable features, the most significant of which is the non-rigorous spatial decomposition scheme. Monte Carlo methods, which allow detailed and accurate modeling of the full geometry and are considered the ''gold standard'' for radiation transport solutions, are playing an ever-increasing role in correcting and/or verifying the deterministic, multi-level spatial decomposition methodology in current practice. However, the prohibitive computational requirements associated with obtaining fully converged, system-wide solutions restrict the role of MC to benchmarking deterministic results at a limited number of state-points for a limited number of relevant quantities. The goal of this research is to change this paradigm by enabling direct use of MC for full-core reactor analyses. The most significant of the many technical challenges that must be overcome are the slow, non-uniform convergence of system-wide MC estimates and the memory requirements associated with detailed solutions throughout a reactor (problems involving hundreds of millions of different material and tally regions due to fuel irradiation, temperature distributions, and the needs associated with multi-physics code coupling). To address these challenges, our research has focused on the development and implementation of (1) a novel hybrid deterministic/MC method for determining high-precision fluxes throughout the problem space in k-eigenvalue problems and (2) an efficient MC domain-decomposition (DD) algorithm that partitions the problem phase space onto multiple processors for massively parallel systems, with statistical uncertainty estimation. The hybrid method development is based on an extension of the FW-CADIS method, which attempts to achieve uniform statistical uncertainty throughout a designated problem space. The MC DD development is being implemented in conjunction with the Denovo deterministic radiation transport package to have direct access to the 3-D, massively parallel discrete-ordinates solver (to support the hybrid method) and the associated parallel routines and structure. This paper describes the hybrid method, its implementation, and initial testing results for a realistic 2-D quarter core pressurized-water reactor model and also describes the MC DD algorithm and its implementation.

21 SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLANTS↗

HBMax: Optimizing Memory Efficiency for Parallel Influence Maximization on Multicore Architectures

The goal of influence maximization is to select k most-influential vertices or seeds in a network, where influence is defined by a given diffusion process. The problem has a number of important applications such as viral marketing, information spread, and epidemic control. Although computing optimal seed set is NP-Hard, due to the submodular nature of the problem efficient approximation algorithms exist. However, even state-of-the-art parallel implementations are limited by a sampling step that incurs large memory footprints. This in turn limits the problem size reach and approximation quality. In this work, we study the memory footprint of the sampling process collecting reverse reachability information in the IMM algorithm over large real-world social networks. We present an adaptive and memory-efficient optimization approach for a state-of-the-art multi-threaded parallel influence maximization algorithm. Our approach,HuffMax, uses a portion of the reverse reachable (RR) sets collected by the algorithm to learn the characteristics of the graph. Then, it compresses the intermediate reverse reachability information with Huffman coding, and queries directly on the compressed data to preserve the memory savings obtained through compression. We also propose an efficient sampling strategy based on the distribution of RR sets, which can further reduce the computation time for typical social networks with long-tail distributions. Considering a NUMA architecture, we scale up our solution on 128-core CPUs and reduce the memory footprint by up to 45.7% with negligible time overhead (or even faster) and without perceivable loss of accuracy.

Chen, Xinyu↗

Parallel Memory-Independent Communication Bounds for SYRK

In this paper, we focus on the parallel communication cost of multiplying a matrix with its transpose, known as a symmetric rank-k update (SYRK). SYRK requires half the computation of general matrix multiplication because of the symmetry of the output matrix. Recent work (Beaumont et al., SPAA '22) has demonstrated that the sequential I/O complexity of SYRK is also a constant factor smaller than that of general matrix multiplication. Inspired by this progress, we establish memory-independent parallel communication lower bounds for SYRK with smaller constants than general matrix multiplication, and we show that these constants are tight by presenting communication-optimal algorithms. The crux of the lower bound proof relies on extending a key geometric inequality to symmetric computations and analytically solving a constrained nonlinear optimization problem. Here, the optimal algorithms use a triangular blocking scheme for parallel distribution of the symmetric output matrix and corresponding computation.

Communication costs↗

Experiments and Analyses of Data Transfers Over Wide-Area Dedicated Connections

Dedicated wide-area network connections are increasingly employed in high-performance computing and big data scenarios. One might expect the performance and dynamics of data transfers over such connections to be easy to analyze due to the lack of competing traffic. However, non-linear transport dynamics and end-system complexities (e.g., multi-core hosts and distributed filesystems) can in fact make analysis surprisingly challenging. We present extensive measurements of memory-to-memory and disk-to-disk file transfers over 10 Gbps physical and emulated connections with 0–366 ms round trip times (RTTs). For memory-to-memory transfers, profiles of both TCP and UDT throughput as a function of RTT show concave and convex regions; large buffer sizes and more parallel flows lead to wider concave regions, which are highly desirable. TCP and UDT both also display complex throughput dynamics, as indicated by their Poincare maps and Lyapunov exponents. For disk-to-disk transfers, we determine that high throughput can be achieved via a combination of parallel I/O threads, parallel network threads, and direct I/O mode. Our measurements also show that Lustre filesystems can be mounted over long-haul connections using LNet routers, although challenges remain in jointly optimizing file I/O and transport method parameters to achieve peak throughput.

Rao, Nageswara S.↗