Search NASA⌕ Search

SEARCH · Search NASA

Results for “distributed-memory architectures”

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.

48 records · Page 3

GPU-acceleration of the ELPA2 distributed eigensolver for dense symmetric and hermitian eigenproblems

The solution of eigenproblems is often a key computational bottleneck that limits the tractable system size of numerical algorithms, among them electronic structure theory in chemistry and in condensed matter physics. Large eigenproblems can easily exceed the capacity of a single compute node, thus must be solved on distributed-memory parallel computers. We here present GPU-oriented optimizations of the ELPA two-stage tridiagonalization eigensolver (ELPA2). On top of cuBLAS-based GPU offloading, we add a CUDA kernel to speed up the back-transformation of eigenvectors, which can be the computationally most expensive part of the two-stage tridiagonalization algorithm. Furthermore, we benchmark the performance of this GPU-accelerated eigensolver on two hybrid CPU–GPU architectures, namely a compute cluster based on Intel Xeon Gold CPUs and NVIDIA Volta GPUs, and the Summit supercomputer based on IBM POWER9 CPUs and NVIDIA Volta GPUs. Consistent with previous benchmarks on CPU-only architectures, the GPU-accelerated two-stage solver exhibits a parallel performance superior to the one-stage counterpart. Finally, we demonstrate the performance of the GPU-accelerated eigensolver developed in this work for routine semi-local KS-DFT calculations comprising thousands of atoms.

97 MATHEMATICS AND COMPUTING↗

ATDM LANL FleCSI: Topology and Execution Framework

FleCSI is a compile-time configurable C++ framework designed to support multi-physics application development. As such, FleCSI attempts to provide a very general set of infrastructure design patterns that can be specialized and extended to suit the needs of a broad variety of solver and data requirements. This means that FleCSI is potentially useful to many different ECP projects. Current support includes multidimensional mesh topology, mesh geometry, and mesh adjacency information, n-dimensional hashed-tree data structures, graph partitioning interfaces, and dependency closures (to identify data dependencies between distributed-memory address spaces). FleCSI introduces a functional programming model with control, execution, and data abstractions that are consistent with state-of-the-art task-based runtimes such as Legion and Charm++. The model also provides support for fine-grained, data-parallel execution with backend support for runtimes such as OpenMP and C++17. The FleCSI abstraction layer provides the developer with insulation from the underlying runtimes, while allowing support for multiple runtime systems, including conventional models like asynchronous MPI. The intent is to give developers a concrete set of user-friendly programming tools that can be used now, while allowing flexibility in choosing runtime implementations and optimizations that can be applied to architectures and runtimes that arise in the future. This project is essential to the ECP Ristra Next-Generation Code project, part of ASC ATDM, because it provides a hierarchically parallel programming model that is consistent with the design of modern system architectures, but which allows for the straightforward expression of algorithmic parallelism in a portably performant manner.

97 MATHEMATICS AND COMPUTING↗

PRIMA-X - Performance Retargeting of Instrumentation, Measurement, and Analysis Technologies for Exascale Computing

The PRIMA-X (Performance Retargeting of Instrumentation, Measurement, and Analysis Technologies for Exascale Computing) project is the successor of the DOE PRIMA (Performance Refactoring of Instrumentation, Measurement, and Analysis Technologies for Petascale Computing) project, which addressed the challenge of creating a core measurement infrastructure that would serve as a common platform for both integrating leading parallel performance systems (notably TAU and Scalasca) and developing next-generation scalable performance tools. The PRIMA-X project shifts the focus away from refactorization of robust performance tools towards a re-targeting of the parallel performance measurement and analysis architecture for extreme scales. The massive concurrency, asynchronous execution dynamics, hardware heterogeneity, and multi-objective prerequisites (performance, power, resilience) that identify exascale systems introduce fundamental constraints on the ability to carry forward existing performance methodologies. In particular, there must be a deemphasis of per-thread observation techniques to significantly reduce the otherwise unsustainable flood of redundant performance data. Instead, it will be necessary to assimilate multi-level resource observations into macroscopic performance views, from which resilient performance metrics can be attributed to the computational features of the application. This requires a scalable framework for node-level and system-wide monitoring and runtime analyses of dynamic performance information. Also, the interest in optimizing parallelism parameters with respect to performance and energy drives the integration of tool capabilities in the exascale environment further. Initially, PRIMA-X was a collaborative project between the University of Oregon (lead institution) and the German Research School for Simulation Sciences (GRS). Because Prof. Wolf, the PI at GRS, accepted a position as full professor at Technische Universitaet Darmstadt (TU Darmstadt) starting February 1st, 2015, the work of GRS was continued at TU Darmstadt. This report covers the work at TU Darmstadt after the transition. The first main accomplishment of TU Darmstadt is the development of different techniques to aggregate performance data on the level of threads. TU Darmstadt evaluated different schemes designed at GRS during the first phase of this project and integrated them into Score-P, a widely used performance measurement framework for HPC applications. The second main accomplishment is a substantial increase of Score-P’s scalability, achieved by improving its internal representation of the underlying system. Third, we extended our performance-modeling tool Extra-P to automatically create models with more than one input parameter (e.g., process count and input size), which was previously impossible. These models allow the performance behavior of an application to be extrapolated to larger machines. Specific applications of the new multi-parameter capability realized in this project include modeling the isoefficiency function of task-based applications and a novel co-design methodology. Further extensions of Extra-P comprise online modeling functionality, automatic refinement of the model search space, and segmented performance modeling. Further contributions include new locality metrics for both shared- and distributed-memory applications, a new way of measuring the impact of inter-application inference on the execution time of an application, and finally an automatic method that selects appropriate OpenMP constructs for the parallelization of sequential code and classifies the affected variables based on their sharing semantics.

97 MATHEMATICS AND COMPUTING↗

Parallel-vector algorithms for particle simulations on shared-memory multiprocessors

Over the last few decades, the computational demands of massive particle-based simulations for both scientific and industrial purposes have been continuously increasing. Hence, considerable efforts are being made to develop parallel computing techniques on various platforms. In such simulations, particles freely move within a given space, and so on a distributed-memory system, load balancing, i.e., assigning an equal number of particles to each processor, is not guaranteed. However, shared-memory systems achieve better load balancing for particle models, but suffer from the intrinsic drawback of memory access competition, particularly during (1) paring of contact candidates from among neighboring particles and (2) force summation for each particle. Here, novel algorithms are proposed to overcome these two problems. For the first problem, the key is a pre-conditioning process during which particle labels are sorted by a cell label in the domain to which the particles belong. Then, a list of contact candidates is constructed by pairing the sorted particle labels. For the latter problem, a table comprising the list indexes of the contact candidate pairs is created and used to sum the contact forces acting on each particle for all contacts according to Newton's third law. With just these methods, memory access competition is avoided without additional redundant procedures. The parallel efficiency and compatibility of these two algorithms were evaluated in discrete element method (DEM) simulations on four types of shared-memory parallel computers: a multicore multiprocessor computer, scalar supercomputer, vector supercomputer, and graphics processing unit. The computational efficiency of a DEM code was found to be drastically improved with our algorithms on all but the scalar supercomputer. Thus, the developed parallel algorithms are useful on shared-memory parallel computers with sufficient memory bandwidth.

97 MATHEMATICS AND COMPUTING↗

MassiveGNN: Efficient Training via Prefetching for Massively Connected Distributed Graphs

Graph Neural Networks (GNN) are indispensable in learning from graph-structured data, yet their rising computational costs, especially on massively connected graphs, pose significant challenges in terms of execution performance. To tackle this, distributed-memory solutions such as partitioning the graph to concurrently train multiple replicas of GNNs are in practice. However, approaches requiring a partitioned graph usually suffer from communication overhead and load imbalance, even under optimal partitioning and communication strategies due to irregularities in the neighborhood minibatch sampling. This paper proposes practical trade-offs for improving the sampling and communication overheads for representation learn- ing on distributed graphs (using popular GraphSAGE architecture) by developing a parameterized prefetch and eviction scheme on top of the state-of-the-art Amazon DistDGL distributed GNN framework, demonstrating about 15–40% improvement in end-to-end training performance on the NERSC Perlmutter supercomputer for various OGB datasets.

Machine Leanring, high performance comptuing, grap↗

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↗

Designing a unified programming model for heterogeneous machines

While high-efficiency machines are increasingly embracing heterogeneous architectures and massive multithreading, contemporary mainstream programming languages reflect a mental model in which processing elements are homogeneous, concurrency is limited, and memory is a flat undifferentiated pool of storage. Moreover, the current state of the art in programming heterogeneous machines tends towards using separate programming models, such as OpenMP and CUDA, for different portions of the machine. Both of these factors make programming emerging heterogeneous machines unnecessarily difficult. We describe the design of the Phalanx programming model, which seeks to provide a unified programming model for heterogeneous machines. It provides constructs for bulk parallelism, synchronization, and data placement which operate across the entire machine. Our prototype implementation is able to launch and coordinate work on both CPU and GPU processors within a single node, and by leveraging the GASNet runtime, is able to run across all the nodes of a distributed-memory machine.

Garland, Michael↗

diBELLA: Distributed Long Read to Long Read Alignment

© 2019 ACM. We present a parallel algorithm and scalable implementation for genome analysis, specifically the problem of finding overlaps and alignments for data from "third generation" long read sequencers [29]. While long sequences of DNA offer enormous advantages for biological analysis and insight, current long read sequencing instruments have high error rates and therefore require different approaches to analysis than their short read counterparts. Our work focuses on an efficient distributed-memory parallelization of an accurate single-node algorithm for overlapping and aligning long reads. We achieve scalability of this irregular algorithm by addressing the competing issues of increasing parallelism, minimizing communication, constraining the memory footprint, and ensuring good load balance. The resulting application, diBELLA, is the first distributed memory overlapper and aligner specifically designed for long reads and parallel scalability. We describe and present analyses for high level design trade-offs and conduct an extensive empirical analysis that compares performance characteristics across state-of-the-art HPC systems as well as a commercial cloud architectures, highlighting the advantages of state-of-the-art network technologies.

Ellis, Marquita↗

PLANC: Parallel Low-rank Approximation with Nonnegativity Constraints

In this work, we consider the problem of low-rank approximation of massive dense nonnegative tensor data, for example, to discover latent patterns in video and imaging applications. As the size of data sets grows, single workstations are hitting bottlenecks in both computation time and available memory. We propose a distributed-memory parallel computing solution to handle massive data sets, loading the input data across the memories of multiple nodes, and performing efficient and scalable parallel algorithms to compute the low-rank approximation. We present a software package called Parallel Low-rank Approximation with Nonnegativity Constraints, which implements our solution and allows for extension in terms of data (dense or sparse, matrices or tensors of any order), algorithm (e.g., from multiplicative updating techniques to alternating direction method of multipliers), and architecture (we exploit GPUs to accelerate the computation in this work). We describe our parallel distributions and algorithms, which are careful to avoid unnecessary communication and computation, show how to extend the software to include new algorithms and/or constraints, and report efficiency and scalability results for both synthetic and real-world data sets.

97 MATHEMATICS AND COMPUTING↗

LDRD final report : massive multithreading applied to national infrastructure and informatics.

Large relational datasets such as national-scale social networks and power grids present different computational challenges than do physical simulations. Sandia's distributed-memory supercomputers are well suited for solving problems concerning the latter, but not the former. The reason is that problems such as pattern recognition and knowledge discovery on large networks are dominated by memory latency and not by computation. Furthermore, most memory requests in these applications are very small, and when the datasets are large, most requests miss the cache. The result is extremely low utilization. We are unlikely to be able to grow out of this problem with conventional architectures. As the power density of microprocessors has approached that of a nuclear reactor in the past two years, we have seen a leveling of Moores Law. Building larger and larger microprocessor-based supercomputers is not a solution for informatics and network infrastructure problems since the additional processors are utilized to only a tiny fraction of their capacity. An alternative solution is to use the paradigm of massive multithreading with a large shared memory. There is only one instance of this paradigm today: the Cray MTA-2. The proposal team has unique experience with and access to this machine. The XMT, which is now being delivered, is a Red Storm machine with up to 8192 multithreaded 'Threadstorm' processors and 128 TB of shared memory. For many years, the XMT will be the only way to address very large graph problems efficiently, and future generations of supercomputers will include multithreaded processors. Roughly 10 MTA processor can process a simple short paths problem in the time taken by the Gordon Bell Prize-nominated distributed memory code on 32,000 processors of Blue Gene/Light. We have developed algorithms and open-source software for the XMT, and have modified that software to run some of these algorithms on other multithreaded platforms such as the Sun Niagara and Opteron multi-core chips.

97 MATHEMATICS AND COMPUTING↗

MASSIVE HYBRID PARALLELISM FOR FULLY IMPLICIT MULTIPHYSICS

As hardware advances continue to modify the supercomputing landscape, traditional scientific software development practices will become more outdated, ineffective, and inefficient. The process of rewriting/retooling existing software for new architectures is a Sisyphean task, and results in substantial hours of development time, effort, and money. Software libraries which provide an abstraction of the resources provided by such architectures are therefore essential if the computational engineering and science communities are to continue to flourish in this modern computing environment. The Multiphysics Object Oriented Simulation Environment (MOOSE) framework enables complex multiphysics analysis tools to be built rapidly by scientists, engineers, and domain specialists, while also allowing them to both take advantage of current HPC architectures, and efficiently prepare for future supercomputer designs. MOOSE employs a hybrid shared-memory and distributed-memory parallel model and provides a complete and consistent interface for creating multiphysics analysis tools. In this paper, a brief discussion of the mathematical algorithms underlying the framework and the internal object-oriented hybrid parallel design are given. Representative massively parallel results from several applications areas are presented, and a brief discussion of future areas of research for the framework are provided.

97 MATHEMATICS AND COMPUTING↗

Massive hybrid parallelism for fully implicit multiphysics

As hardware advances continue to modify the supercomputing landscape, traditional scientific software development practices will become more outdated, ineffective, and inefficient. The process of rewriting/retooling existing software for new architectures is a Sisyphean task, and results in substantial hours of development time, effort, and money. Software libraries which provide an abstraction of the resources provided by such architectures are therefore essential if the computational engineering and science communities are to continue to flourish in this modern computing environment. The Multiphysics Object Oriented Simulation Environment (MOOSE) framework enables complex multiphysics analysis tools to be built rapidly by scientists, engineers, and domain specialists, while also allowing them to both take advantage of current HPC architectures, and efficiently prepare for future supercomputer designs. MOOSE employs a hybrid shared-memory and distributed-memory parallel model and provides a complete and consistent interface for creating multiphysics analysis tools. In this paper, a brief discussion of the mathematical algorithms underlying the framework and the internal object-oriented hybrid parallel design are given. Representative massively parallel results from several applications areas are presented, and a brief discussion of future areas of research for the framework are provided. (authors)

97 MATHEMATICS AND COMPUTING↗