Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Parallel Supercomputers”

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 649 records · Page 36

Parallel algorithms for finding connected components using linear algebra

Finding connected components is one of the most widely used operations on a graph. Optimal serial algorithms for the problem have been known for half a century, and many competing parallel algorithms have been proposed over the last several decades under various different models of parallel computation. This paper presents a class of parallel connected-component algorithms designed using linear-algebraic primitives. These algorithms are based on a PRAM algorithm by Shiloach and Vishkin and can be designed using standard GraphBLAS operations. Here, we demonstrate two algorithms of this class, one named LACC for Linear Algebraic Connected Components, and the other named FastSV which can be regarded as LACC’s simplification. With the support of the highly-scalable Combinatorial BLAS library, LACC and FastSV outperform the previous state-of-the-art algorithm by a factor of up to 12x for small to medium scale graphs. For large graphs with more than 50B edges, LACC and FastSV scale to 4K nodes (262K cores) of a Cray XC40 supercomputer and outperform previous algorithms by a significant margin. This remarkable performance is accomplished by (1) exploiting sparsity that was not present in the original PRAM algorithm formulation, (2) using high-performance primitives of Combinatorial BLAS, and (3) identifying hot spots and optimizing them away by exploiting algorithmic insights.

97 MATHEMATICS AND COMPUTING↗

Mesoscopic Modeling and Rapid Simulation of Incremental Changes in Epidemic Scenarios on GPUs

In simulation-based studies and analyses of epidemics, a major challenge lies in resolving the conflict between fidelity of models and the speed of their simulation. Another related challenge arises in dealing with the large number of what–if scenarios that need to be explored. Here, we describe new computational methods that together provide an approach to dealing with both challenges. A mesoscopic modeling approach is described that strikes a middle ground between macroscopic models based on coupled differential equations and microscopic models built on fine-grained behaviors at the individual entity level. The mesoscopic approach offers the ability to incorporate complex compositions of multiple layers of dynamics even while retaining the potential for aggregate behaviors at varying levels. It also is an excellent match to the accelerator-based architectures of modern computing platforms in which graphical processing units (GPUs) can be exploited for fast simulation via the parallel execution mode of single instruction multiple thread (SIMT). The challenge of simulating a large number of scenarios is addressed via a method of sharing model state and computation across a tree of what–if scenarios that are localized, incremental changes to a large base simulation. A combination of the mesoscopic modeling approach and the incremental what–if scenario tree evaluation has been implemented in the software on modern GPUs. Synthetic simulation scenarios are presented to demonstrate the computational characteristics of our approach. Results from the experiments with large population data, including USA, UK, and India, illustrate the modeling methodology and computational performance on thousands of synthetically generated what–if scenarios. Execution of our implementation scaled to 8192 GPUs of supercomputing platforms demonstrates the ability to rapidly evaluate what–if scenarios several orders of magnitude faster than the conventional methods.

97 MATHEMATICS AND COMPUTING↗

A large-scale study of soft-errors on GPUs in the field

Parallelism provided by the GPU architecture has enabled domain scientists to simulate physical phenomena at a much faster rate and finer granularity than what was previously possible by CPU-based large-scale clusters. Architecture researchers have been investigating reliability characteristics of GPUs and innovating techniques to increase the reliability of these emerging computing devices. Such efforts are often guided by technology projections and simplistic scientific kernels, and performed using architectural simulators and modeling tools. Lack of large-scale field data impedes the effectiveness of such efforts. This study attempts to bridge this gap by presenting a large-scale field data analysis of GPU reliability. We characterize and quantify different kinds of soft-errors on the Titan supercomputer's GPU nodes. Our study uncovers several interesting and previously unknown insights about the characteristics and impact of soft-errors.

Nie, Bin↗

GSoFa: Scalable Sparse Symbolic LU Factorization on GPUs

Decomposing a matrix $\mathbf {A}$ into a lower matrix $\mathbf {L}$ and an upper matrix $\mathbf {U}$, which is also known as LU decomposition, is an essential operation in numerical linear algebra. For a sparse matrix, LU decomposition often introduces more nonzero entries in the $\mathbf {L}$ and $\mathbf {U}$ factors than in the original matrix. A symbolic factorization step is needed to identify the nonzero structures of $\mathbf {L}$ and $\mathbf {U}$ matrices. Attracted by the enormous potentials of the Graphics Processing Units (GPUs), an array of efforts have surged to deploy various LU factorization steps except for the symbolic factorization, to the best of our knowledge, on GPUs. This article introduces gSoFa, the first GPU-based symbolic factorization design with the following three optimizations to enable scalable LU symbolic factorization for nonsymmetric pattern sparse matrices on GPUs. First, here we introduce a novel fine-grained parallel symbolic factorization algorithm that is well suited for the Single Instruction Multiple Thread (SIMT) architecture of GPUs. Second, we tailor supernode detection into a SIMT friendly process and strive to balance the workload, minimize the communication and saturate the GPU computing resources during supernode detection. Third, we introduce a three-pronged optimization to reduce the excessive space consumption problem faced by multi-source concurrent symbolic factorization. Taken together, gSoFa achieves up to 31× speedup from 1 to 44 Summit nodes (6 to 264 GPUs) and outperforms the state-of-the-art CPU project, on average, by 5×. Notably, gSoFa also achieves up to 47 percent of the peak memory throughput of a V100 GPU in the Summit Supercomputer.

97 MATHEMATICS AND COMPUTING↗

Exploring flexible communications for streamlining DNN ensemble training pipelines

Parallel training of a Deep Neural Network (DNN) ensemble on a cluster of nodes is a common practice to train multiple models in order to construct a model with a higher prediction accuracy, or to quickly tune the parameters of a training model. Existing ensemble training pipelines perform a great deal of redundant operations, resulting in unnecessary CPU usage, or even poor pipeline performance. In order to remove these redundancies, we need pipelines with more communication flexibility than existing DNN frameworks can provide. This project investigates a series of designs to improve pipeline flexibility and adaptivity, while also increasing performance. We implement our designs using Tensorflow with Horovod, and test it using several large DNNs in a large scale GPU cluster, the Titan supercomputer at Oak Ridge National Lab. Our results show that with the new flexible communication schemes, the CPU time spent during training is reduced by 2-11X. Furthermore, our implementation can achieve up to 10X speedups when CPU core limits are imposed. Our best pipeline also reduces the average power draw of the ensemble training process by 5--16% when compared to the baseline.

Pittman, Randall↗

Collaboration Portal for Petascale Simulations

The emergence of leadership class computing is creating a tsunami of data from petascale simulations. Results are typically analyzed by dozens of scientists. In order for the scientist to digest the vast amount of data being produced from the simulations and auxiliary programs, it is critical to automate the effort to manage, analyze, visualize, and share this data. One aspect of this is leveraging of their collective knowledge and experiences through a scientific social network. This can be archived through a combination of parallel back-end services, provenance capturing, and an easy to use front-end tool. 'eSimMon', is one such tool we developed as part of the Scientific Discovery through Advanced Computing (SciDAC) program. In this paper we describe eSimMon, discuss its ease of use, its efficiency, and its ability to accelerate scientific discovery through advanced computing.

97 MATHEMATICS AND COMPUTING↗

Linux Kernel Co-Scheduling For Bulk Synchronous Parallel Applications

This paper describes a kernel scheduling algorithm that is based on co-scheduling principles and that is intended for parallel applications running on 1000 cores or more where inter-node scalability is key. Experimental results for a Linux implementation on a Cray XT5 machine are presented.1 The results indicate that Linux is a suitable operating system for this new scheduling scheme, and that this design provides a dramatic improvement in scaling performance for synchronizing collective operations at scale.

99 GENERAL AND MISCELLANEOUS↗

Scalable Asynchronous Domain Decomposition Solvers

We discuss how parallel implementations of linear iterative solvers generally alternate between phases of data exchange and phases of local computation. Increasingly large problem sizes and more heterogeneous compute architectures make load balancing and the design of low latency network interconnects that are able to satisfy the communication requirements of linear solvers very challenging tasks. In particular, global communication patterns such as inner products become increasingly limiting at scale. We explore the use of asynchronous communication based on one-sided Message Passing Interface primitives in the context of domain decomposition solvers. In particular, a scalable asynchronous two-level Schwarz method is presented. We discuss practical issues encountered in the development of a scalable solver and show experimental results obtained on a state-of-the-art supercomputer system that illustrate the benefits of asynchronous solvers in load balanced as well as load imbalanced scenarios. Using the novel method, we can observe speedups of up to four times over its classical synchronous equivalent.

97 MATHEMATICS AND COMPUTING↗

Distributed Halide

Many image processing tasks are naturally expressed as a pipeline of small computational kernels known as stencils. Halide is a popular domain-specific language and compiler designed to implement image processing algorithms. Halide uses simple language constructs to express what to compute and a separate scheduling co-language for expressing when and where to perform the computation. This approach has demonstrated performance comparable to or better than hand-optimized code. Until now, however, Halide has been restricted to parallel shared memory execution, limiting its performance for memory-bandwidth-bound pipelines or large-scale image processing tasks. We present an extension to Halide to support distributed-memory parallel execution of complex stencil pipelines. These extensions compose with the existing scheduling constructs in Halide, allowing expression of complex computation and communication strategies. Existing Halide applications can be distributed with minimal changes, allowing programmers to explore the tradeoff between recomputation and communication with little effort. Approximately 10 new of lines code are needed even for a 200 line, 99 stage application. On nine image processing benchmarks, our extensions give up to a 1.4× speedup on a single node over regular multithreaded execution with the same number of cores, by mitigating the effects of non-uniform memory access. The distributed benchmarks achieve up to 18× speedup on a 16 node testing machine and up to 57× speedup on 64 nodes of the NERSC Cori supercomputer.

97 MATHEMATICS AND COMPUTING↗

Library for Evolutionary Algorithms in Python (LEAP)

There are generally three types of scientific software users: users that solve problems using existing science software tools, researchers that explore new approaches by extending existing code, and educators that teach students scientific concepts. Python is a general-purpose programming language that is accessible to beginners, such as students, but also as a language that has a rich scientific programming ecosystem that facilitates writing research software. Additionally, as high-performance computing (HPC) resources become more readily available, software support for parallel processing becomes more relevant to scientific software.There currently are no Python-based evolutionary computation frameworks that support all three types of scientific software users. Moreover, some support synchronous concurrent fitness evaluation that do not efficiently use HPC resources. We pose here a new Python-based EC framework that uses an established generalized unified approach to EA concepts to provide an easy to use toolkit for users wishing to use an EA to solve a problem, for researchers to implement novel approaches, and for providing a low-bar to entry to EA concepts for students. Additionally, this toolkit provides a scalable asynchronous fitness evaluation implementation friendly to HPC that has been vetted on hardware ranging from laptops to the world’s fastest supercomputer, Summit.

Coletti, Mark↗

Parallel-in-Time Simulation of Lindblad's Equation

Constructing fast quantum logic gates is critical to building a scalable quantum computer. We consider a qudit, a quantum version of a bit that can take an arbitrary number of states, coupled with a cavity. In this project, we wish to force the qudit to reach the 0-state, for any possible initial state. The coupled system changes in time according to Lindblad’s equation, an ordinary differential equation on the density matrix of the quantum system. Lindblad’s equation contains some parameters that we can control, so-called control functions. We seek control functions which force the qudit to the 0-state within 2 microseconds, which is much faster than what is currently done in practice. The search method is gradient descent, a numerical optimization method that uses gradient information to iteratively improve the control parameters. My contribution to this project is an attempt to speed up the computation of the gradient. It currently takes about 40 seconds to compute the gradient which involves solving a set of ODEs sequentially. Current supercomputers have thousands of cores, but sequential computations can only make use of 1 core at a time. We wish to divide up the work better, so that we can use many more cores at once. To this end, we have implemented the Multigrid Reduction in Time (MGRIT) algorithm. We perform a systematic parameter search on how to best apply this algorithm. Results indicate a 25 percent speed up for solving Lindblad’s equation and determining how close the final state is the 0-state.

97 MATHEMATICS AND COMPUTING↗

Pseudonymization at Scale: OLCF’s Summit Usage Data Case Study

The analysis of vast amounts of data and the processing of complex computational jobs have traditionally relied upon high performance computing (HPC) systems, which offer reliable and efficient management of large-scale computational and data resources. Understanding these analyses’ needs is paramount for designing solutions that can lead to better science, and similarly, understanding the characteristics of the user behavior on those systems is important for improving user experiences on HPC systems. A common approach to gathering data about user behavior is to extract workload characteristics from system log data available only to system administrators. Recently at Oak Ridge Leadership Computing Facility (OLCF), however, we unveiled user behavior about the Summit supercomputer by collecting data from a user’s point of view with ordinary Unix commands.In this paper, we discuss the process, challenges, and lessons learned while preparing this dataset for publication and submission to an open data challenge. The original dataset contains personal identifiable information (PII) about the users of OLCF which needed be masked prior to publication, and we determined that anonymization, which scrubs PII completely, destroyed too much of the structure of the data to be interesting for the data challenge. We instead chose to pseudonymize the dataset, which reduced the linkability of the dataset to the users’ identities. Pseudonymization is significantly more computationally expensive than anonymization, and the size of our dataset, which is approximately 175 million lines of raw text, necessitated the development of a parallelized workflow that could be reused on different HPC machines. We demonstrate the scaling behavior of the workflow on two leadership class HPC systems at OLCF, and we show that we were able to bring the overall makespan time from an impractical 20+ hours on a single node down to around 2 hours. As a result of this work, we release the entire pseudonymized dataset and make the workflows and source code publicly available.

Maheshwari, Ketan↗

Unique Methodologies for Nano/Micro Manufacturing Job Training Via Desktop Supercomputer Modeling and Simulation

This project establishes an initiative in high speed (Teraflop)/large-memory desktop supercomputing for modeling and simulation of dynamic processes important for energy and industrial applications. It provides a training ground for employment of current students in an emerging field with skills necessary to access the large supercomputing systems now present at DOE laboratories. It also provides a foundation for NIU faculty to quantum leap beyond their current small cluster facilities. The funding extends faculty and student capability to a new level of analytic skills with concomitant publication avenues. The components of the Hewlett Packard computer obtained by the DOE funds create a hybrid combination of a Graphics Processing System (12 GPU/Teraflops) and a Beowulf CPU system (144 CPU), the first expandable via the NIU GAEA system to ~60 Teraflops integrated with a 720 CPU Beowulf system. The software is based on access to the NVIDIA/CUDA library and the ability through MATLAB multiple licenses to create additional local programs. A number of existing programs are being transferred to the CPU Beowulf Cluster. Since the expertise necessary to create the parallel processing applications has recently been obtained at NIU, this effort for software development is in an early stage. The educational program has been initiated via formal tutorials and classroom curricula designed for the coming year. Specifically, the cost focus was on hardware acquisitions and appointment of graduate students for a wide range of applications in engineering, physics and computer science.

Kimball, Clyde↗

Prometheus: Scalable and Accurate Emulation of Task-Based Applications on Many-Core Systems.

Modeling the performance of non-deterministic parallel applications on future many-core systems requires the development of novel simulation and emulation techniques and tools. We present “Prometheus”, a fast, accurate and modular emulation framework for task-based applications. By raising the level of abstraction and focusing on runtime synchronization, Prometheus can accurately predict applications’ performance on very large many-core systems. We validate our emulation framework against two real platforms (AMD Interlagos and Intel MIC) and report error rates generally below 4%. We, then, evaluate Prometheus’ performance and scalability: our results show that Prometheus can emulate a task-based application on a system with 512K cores in 11.5 hours. We present two test cases that show how Prometheus can be used to study the performance and behavior of systems that present some of the characteristics expected from exascale supercomputer nodes, such as active power management and processors with a high number of cores but reduced cache per core.

exascale systems↗

HipBone: A performance-portable graphics processing unit-accelerated C++ version of the NekBone benchmark

We present hipBone, an open-source performance-portable proxy application for the Nek5000 (and NekRS) computational fluid dynamics applications. HipBone is a fully GPU-accelerated C++ implementation of the original NekBone CPU proxy application with several novel algorithmic and implementation improvements which optimize its performance on modern fine-grain parallel GPU accelerators. Our optimizations include a conversion to store the degrees of freedom of the problem in assembled form in order to reduce the amount of data moved during the main iteration and a portable implementation of the main Poisson operator kernel. We demonstrate near-roofline performance of the operator kernel on three different modern GPU accelerators from two different vendors. We present a novel algorithm for splitting the application of the Poisson operator on GPUs which aggressively hides MPI communication required for both halo exchange and assembly. Our implementation of nearest-neighbor MPI communication then leverages several different routing algorithms and GPU-Direct RDMA capabilities, when available, which improves scalability of the benchmark. We demonstrate the performance of hipBone on three different clusters housed at Oak Ridge National Laboratory, namely, the Summit supercomputer and the Frontier early-access clusters, Spock and Crusher. Our tests demonstrate both portability across different clusters and very good scaling efficiency, especially on large problems.

Computer Science↗

Scaling Resolution of Gigapixel Whole Slide Images Using Spatial Decomposition on Convolutional Neural Networks

Gigapixel images are prevalent in scientific domains ranging from remote sensing, and satellite imagery to microscopy, etc. However, training a deep learning model at the natural resolution of those images has been a challenge in terms of both, overcoming the resource limit (e.g. HBM memory constraints), as well as scaling up to a large number of GPUs. In this paper, we trained Residual neural Networks (ResNet) on 22,528 x 22,528-pixel size images using a distributed spatial decomposition method on 2,304 GPUs on the Summit Supercomputer. We applied our method on a Whole Slide Imaging (WSI) dataset from The Cancer Genome Atlas (TCGA) database. WSI images can be in the size of 100,000 x 100,000 pixels or even larger, and in this work we studied the effect of image resolution on a classification task, while achieving state-of-the-art AUC scores. Moreover, our approach doesn't need pixel-level labels, since we're avoiding patching from the WSI images completely, while adding the capability of training arbitrary large-size images. This is achieved through a distributed spatial decomposition method, by leveraging the non-block fat-tree interconnect network of the Summit architecture, which enabled GPU-to-GPU direct communication. Finally, detailed performance analysis results are shown, as well as a comparison with a data-parallel approach when possible.

Tsaris, Aristeidis (aris)↗

A fast, dense Chebyshev solver for electronic structure on GPUs

Matrix diagonalization is almost always involved in computing the density matrix needed in quantum chemistry calculations. In the case of modest matrix sizes (≲4000), performance of traditional dense diagonalization algorithms on modern GPUs is underwhelming compared to the peak performance of these devices. This motivates the exploration of alternative algorithms better suited to these types of architectures. We newly derive, and present in detail, an existing Chebyshev expansion algorithm whose number of required matrix multiplications scales with the square root of the number of terms in the expansion. Focusing on dense matrices of modest size, our implementation on GPUs results in large speed ups when compared to diagonalization. Additionally, we improve upon this existing method by capitalizing on the inherent task parallelism and concurrency in the algorithm. Furthermore, this improvement is implemented on GPUs by using CUDA and HIP streams via the MAGMA library and leads to a significant speed up over the serial-only approach for smaller (≲1000) matrix sizes. Finally, we apply our technique to a model system with a high density of states around the Fermi level, which typically presents significant challenges.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗