Dataflow System: Is It A Viable Scalable Graph Analysis Engine?
Abstract not provided
SEARCH · Search NASA
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.
Abstract not provided
We introduce Σ-Attention, a transformer-based operator-learning framework for approximating the self-energy operator of strongly correlated electronic systems. By creating a batched dataset that combines results from three complementary approaches, i.e., many-body perturbation theory, strong-coupling expansion, and exact diagonalization, each effective in specific parameter regimes, Σ-Attention is applied to learn an accurate approximation for the self-energy operator that is valid across a wide range of parameter regimes. This hybrid strategy leverages the strengths of existing methods while relying on the transformer's ability to generalize beyond individual limitations. More importantly, the scalability of the transformer architecture allows the learned self-energy to be extended to systems with larger sizes, leading to much improved computational scaling. Using the one-dimensional Hubbard model, we demonstrate that Σ-Attention can accurately predict the Matsubara Green's function of large systems with a wide range of coupling strength. Our framework offers a promising and scalable pathway for studying strongly correlated systems with many possible generalizations.
This project developed a generic and optimized set of core data analytics functions. These functions organically consolidate a broad constellation of high performance analytical pipelines. As the architectures of emerging HPC systems become inherently heterogeneous, there is a need to design algorithms for data analysis kernels accelerated on hybrid multi-node, multi-core HPC architectures comprised of a mix of CPUs, GPUs, and SSDs. Furthermore, the power-aware trend drives the advances in our performance-energy tradeoff analysis framework which enables our data analysis kernels algorithms and software to be parameterized so that users can choose the right power-performance optimizations.
This work introduces structure preserving hierarchical decompositions for sampling Gaussian random fields (GRFs) within the context of multilevel Bayesian inference in high-dimensional space. Existing scalable hierarchical sampling methods, such as those based on stochastic partial differential equations (SPDEs), often reduce the dimensionality of the sample space at the cost of accuracy of inference. Other approaches, such that those based on Karhunen-Loève (KL) expansions, offer sample space dimensionality reduction but sacrifice GRF representation accuracy and ergodicity of the Markov chain Monte Carlo (MCMC) sampler and are computationally expensive for high-dimensional problems. The proposed method integrates the dimensionality reduction capabilities of KL expansions with the scalability of SPDE-based sampling, thereby providing a robust, unified framework for high-dimensional uncertainty quantification (UQ) that is scalable and accurate, preserves ergodicity, and offers dimensionality reduction of the sample space. The hierarchy in our multilevel algorithm is derived from the geometric multigrid hierarchy. By constructing a hierarchical decomposition that maintains the covariance structure across the levels in the hierarchy, the approach enables efficient coarse-to-fine sampling while ensuring that all samples are drawn from the desired distribution. The effectiveness of the proposed method is demonstrated on a benchmark subsurface flow problem, demonstrating its effectiveness in improving computational efficiency and statistical accuracy. Furthermore, our proposed technique is more efficient and accurate and displays better convergence properties than existing methods for high-dimensional Bayesian inference problems.
This study explores the feasibility of leveraging quantum machine learning (QML) to analyze NASA Earth Observational (EO) data for climate change research, with a particular focus on the phenomenon of ”crop frosting” which has become more prevalent due to climate change. We implemented and evaluated two QML models, the Variational Quantum Classifier (VQC) and Quantum Support Vector Classifier (QSVC), in both simulated and real quantum computing environments using a 127 qubit IBM quantum processor. Our study emphasizes the scientific rigor in comparing these quantum models with a classical Support Vector Machine (SVM) classifier, highlighting their performance in processing climate data. The results offer valuable insights into the potential scientific advantages, limitations, and scalability of QML for analyzing EO datasets, thus paving the way for more advanced climate modeling and predictive analytics using quantum computing. We showcased how Environmental Interaction Knowledge Graphs (EIKGs) and Digital Twins (DTs) can be integrated into this study. This research underscores the transformative potential of Classical and QML leveraging KGs and DT to address the multifaceted challenges posed by climate change.
The use of clouds to execute high-performance computing (HPC) applications has greatly increased recently. Clouds provide several potential advantages over traditional supercomputers and in-house clusters. The most popular cloud is currently Amazon EC2, which provides fixed-cost and variable-cost, auction-based options. The auction market trades lower cost for potential interruptions that necessitate checkpointing; if the market price exceeds the bid price, a node is taken away from the user without warning. We explore techniques to maximize performance per dollar given a time constraint within which an application must complete. Specifically, we design and implement multiple techniques to reduce expected cost by exploiting redundancy in the EC2 auction market. We then design an adaptive algorithm that selects a scheduling algorithm and determines the bid price. We show that our adaptive algorithm executes programs up to seven times cheaper than using the on-demand market and up to 44 percent cheaper than the best non-redundant, auction-market algorithm. We extend our adaptive algorithm to incorporate application scalability characteristics for further cost savings. In conclusion, we show that the adaptive algorithm informed with scalability characteristics of applications achieves up to 56 percent cost savings compared to the expected cost for the base adaptive algorithm run at a fixed, user-defined scale.
This constitutes the final report for “Dynamic Non-Hierarchical File Systems for Exascale Storage”. The ultimate goal of this project was to improve data management in scientific computing and high-end computing (HEC) applications, and to achieve this goal we proposed: to develop the first, HEC-targeted, file system featuring rich metadata and provenance collection, extreme scalability, and future storage hardware integration as core design goals, and to evaluate and develop a flexible non-hierarchical file system interface suitable for providing more powerful and intuitive data management interfaces to HEC and scientific computing users. Data management is swiftly becoming a serious problem in the scientific community – while copious amounts of data are good for obtaining results, finding the right data is often daunting and sometimes impossible. Scientists participating in a Department of Energy workshop noted that most of their time was spent “...finding, processing, organizing, and moving data and it’s going to get much worse”. Scientists should not be forced to become data mining experts in order to retrieve the data they want, nor should they be expected to remember the naming convention they used several years ago for a set of experiments they now wish to revisit. Ideally, locating the data you need would be as easy as browsing the web. Unfortunately, existing data management approaches are usually based on hierarchical naming, a 40 year-old technology designed to manage thousands of files, not exabytes of data. Today’s systems do not take advantage of the rich array of metadata that current high-end computing (HEC) file systems can gather, including content-based metadata and provenance1 information. As a result, current metadata search approaches are typically ad hoc and often work by providing a parallel management system to the “main” file system, as is done in Linux (the locate utility), personal computers, and enterprise search appliances. These search applications are often optimized for a single file system, making it difficult to move files and their metadata between file systems. Users have tried to solve this problem in several ways, including the use of separate databases to index file properties, the encoding of file properties into file names, and separately gathering and managing provenance data, but none of these approaches has worked well, either due to limited usefulness or scalability, or both. Our research addressed several key issues: High-performance, real-time metadata harvesting: extracting important attributes from files dynamically and immediately updating indexes used to improve search; Transparent, automatic, and secure provenance capture: recording the data inputs and processing steps used in the production of each file in the system; Scalable indexing: indexes that are optimized for integration with the file system; Dynamic file system structure: our approach provides dynamic directories similar to those in semantic file systems, but these are the native organization rather than a feature grafted onto a conventional system. In addition to these goals, our research effort will include evaluating the impact of new storage technologies on the file system design and performance. In particular, the indexing and metadata harvesting functions can potentially benefit from the performance improvements promised by new storage class memories.
Plug-and-play priors (PnP) is a broadly applicable methodology for solving inverse problems by exploiting statistical priors specified as denoisers. Recent work has reported the state-of-the-art performance of PnP algorithms using pre-trained deep neural nets as denoisers in a number of imaging applications. However, current PnP algorithms are impractical in large-scale settings due to their heavy computational and memory requirements. This work addresses this issue by proposing an incremental variant of the widely used PnP-ADMM algorithm, making it scalable to problems involving a large number measurements. Here, we theoretically analyze the convergence of the algorithm under a set of explicit assumptions, extending recent theoretical results in the area. Additionally, we show the effectiveness of our algorithm with nonsmooth data-fidelity terms and deep neural net priors, its fast convergence compared to existing PnP algorithms, and its scalability in terms of speed and memory.
This paper presents a method to perform efficient and automated Overset Grid Assembly (OGA) on a system of overlapping unstructured meshes in a parallel computing environment where all meshes are partitioned into multiple mesh-blocks and processed on multiple cores. The main task of the overset grid assembler is to identify, in parallel, among all points in the overlapping mesh system, at which points the flow solution should be computed (field points), interpolated (receptor points), or ignored (hole points). Point containment search or donor search, an algorithm to efficiently determine the cell that contains a given point, is the core procedure necessary for accomplishing this task. Donor search is particularly challenging for partitioned unstructured meshes because of the complex irregular boundaries that are often created during partitioning. Another challenge arises because of the large variation in the type of mesh-block overlap and the resulting large load imbalance on multiple processors. Desirable traits for the grid assembly method are efficiency (requiring only a small fraction of the solver time), robustness (correct identification of all point types), and full automation (no user input required other than the mesh system). Additionally, the method should be scalable, which is an important challenge due to the inherent load imbalance. This paper describes a fully-automated grid assembly method, which can use two different donor search algorithms. One is based on the use of auxiliary grids and Exact Inverse Maps (EIM), and the other is based on the use of Alternating Digital Trees (ADT). The EIM method is demonstrated to be more efficient than the ADT method, while retaining robustness. An adaptive load re-balance algorithm is also designed and implemented, which considerably improves the scalability of the method.
PIConGPU, Particle In Cell on GPUs, is an open source simulations framework for plasma and laser-plasma physics used to develop advanced particle accelerators for radiation therapy of cancer, high energy physics and photon science. While PIConGPU has been optimized for at least 5 years to run well on NVIDIA GPU-based clusters, there has been limited exploration by the development team of potential scalability bottlenecks using recently updated and new tools including NVIDIA’s NVProf tool and the brand-new NVIDIA NSight Suite (Systems and Compute) tools. PIConGPU is a highly optimized application that runs production jobs at scale on a system Oak Ridge Leadership Facility’s (OLCF) Summit supercomputer (using the full machine at 4600 nodes; at 98% of GPU utilization on all ~28000 NVIDIA Volta GPUs). PIConGPU has been selected as one of the the eight applications for OLCF’s coveted Center for Accelerated Application Readiness (CAAR) program aimed at the facility’s Frontier supercomputer (OLCF’s first exascale system to launch in 2021), to partner with our vendors (primary vendors: AMD and Cray/HPE) ensuring that Frontier will be able to perform large-scale science when it opens to users in 2022. To this effect, performance engineers on the PIConGPU team wanted to dive deep into the application to understand at the finest granularity, which portions of the code could be further optimized to exploit the hardware on Summit at it’s maximum potential and also to elucidate which key kernels should be tracked and optimized for the CAAR effort to port this code to Frontier. Any bottlenecks that are observed via performance profiling on Summit are likely to also impact scalability on the Frontier-dev system and the Frontier Early Access (EA) system. Additionally, the engineers wanted to take a closer look at the newest NVIDIA profiling tools which allows us to identify the most useful features on these tools and will provide an opportunity to compare it to new AMD and Cray’s performance analysis tool releases and provide feedback to our vendor partners on what features are most important and mission critical for CAAR efforts. The primary goal of this report is to focus on the evaluation of PIConGPU’s most time-intensive kernels using NVProf and NSight Suite. Three kernels, Current Deposition (also known as Compute Current), Particle Push (Move and Mark), and Shift Particles are known to be some of the most time-consuming kernels in PIConGPU. The Current xi Deposition kernel and Particle Push kernel both set up the particle attributes for running any physics simulation with PIConGPU, so it is crucial to improve the performance of these two kernels. In this report, we measure single GPU metrics for the three kernels, offer high level takeaways from the conducted analysis, and compare the profiling data from NSight Compute to that of NVProf. This analysis was performed using a grid size of 240 x 272 x 224, and 10 time steps with the Mid-November Figure of Merit (FOM) run setup. The Traveling Wave Electron Acceleration (TWEAC) science case used in this run is a representative science case for PIConGPU. This execution can also be used for baseline analysis on AMD MI50/ MI60 systems. As of the time of writing, the PIConGPU application has limited use for features of NSight Systems, so this report will mainly focus on insights garnered from NSight Compute. For this analysis, we run the “full” metric set available in NSight Compute version 2020.1.2 and use NSight Systems version 2020.3.1 to generate the application timeline.
The Project 8 collaboration aims to measure the absolute neutrino mass or improve on the current limit by measuring the tritium beta decay electron spectrum. We present the current distributed computing model for the Project 8 experiment. Project 8 is in its second phase of data taking with a near continuous data rate of 1Gbps. The current computing model uses DIRAC (Distributed Infrastructure with Remote Agent Control) for its workflow and data management. A detailed meta-data assignment using the DIRAC File Catalog is used to automate raw data transfers and subsequent stages of data processing. The DIRAC system is deployed on containers managed using a Kubernetes cluster to provide a scalable infrastructure. A modified DIRAC Site Director provides the ability to submit jobs using Singularity on opportunistic High-Performance Computing (HPC) sites.
Economic competitiveness and national security depend increasingly on the insightful analysis of large data sets. The diversity of real-world data sources and analytic workflows impose challenging hardware and software requirements for parallel graph platforms. The irregular nature of graph methods is not supported well by the deep memory hierarchies of conventional distributed systems, requiring new processor and runtime system designs to tolerate memory and synchronization latencies. Moreover, the efficiency of relational table operations and matrix computations are not attainable when data is stored in common graph data structures. In this paper, we present HAGGLE, a high-performance, scalable data analytics platform. The platform’s hybrid data model supports a variety of distributed, thread-safe data structures, parallel programming constructs, and persistent and streaming data. An abstract runtime layer enables us to map the stack to conventional, distributed computer systems with accelerators. The runtime uses multithreading, active messages, and data aggregation to hide memory and synchronization latencies on large-scale systems.
A directional implicit unstructured agglomeration multigrid solver is ported to shared and distributed memory massively parallel machines using the explicit domain-decomposition and message-passing approach. Because the algorithm operates on local implicit lines in the unstructured mesh, special care is required in partitioning the problem for parallel computing. A weighted partitioning strategy is described which avoids breaking the implicit lines across processor boundaries, while incurring minimal additional communication overhead. Good scalability is demonstrated on a 128 processor SGI Origin 2000 machine and on a 512 processor CRAY T3E machine for reasonably fine grids. The feasibility of performing large-scale unstructured grid calculations with the parallel multigrid algorithm is demonstrated by computing the flow over a partial-span flap wing high-lift geometry on a highly resolved grid of 13.5 million points in approximately 4 hours of wall clock time on the CRAY T3E.
The CMS Global Pool, based on HTCondor and glideinWMS, is the main computing resource provisioning system for all CMS workflows, including analysis, Monte Carlo production, and detector data reprocessing activities. The total resources at Tier-1 and Tier-2 grid sites pledged to CMS exceed 100,000 CPU cores, while another 50,000 to 100,000 CPU cores are available opportunistically, pushing the needs of the Global Pool to higher scales each year. These resources are becoming more diverse in their accessibility and configuration over time. Furthermore, the challenge of stably running at higher and higher scales while introducing new modes of operation such as multi-core pilots, as well as the chaotic nature of physics analysis workflows, places huge strains on the submission infrastructure. This paper details some of the most important challenges to scalability and stability that the CMS Global Pool has faced since the beginning of the LHC Run II and how they were overcome.
Graph algorithms are challenging to parallelize when high performance and scalability are primary goals. Low concurrency, poor data locality, irregular access pattern, and high data access to computation ratio are among the chief reasons for the challenge. The performance implication of these features is exasperated on distributed memory machines. More success is being achieved on shared-memory, multi-core architectures supporting multithreading. We consider a prototypical graph problem, coloring, and show how a greedy algorithm for solving it can be e*ectively parallelized on multithreaded architectures. We present in particular two di*erent parallel algorithms. The first relies on speculation and iteration, and is suitable for any shared-memory, multithreaded system. The second uses data ow principles and is targeted at the massively multithreaded Cray XMT system. We benchmark the algorithms on three di*erent platforms and demonstrate scalable runtime performance. In terms of quality of solution, both algorithms use nearly the same number of colors as the serial algorithm.
Quantum information science (QIS) holds promise for revolutionizing computation, communication, and sensing. Of course, realizing this promise requires material platforms that can host quantum bits with pre-assigned characteristics that are uniquely attuned to such functionalities. For example, quantum bits for computation should be as immune as possible to any external perturbation, controllable on short time scales and scalable. In contrast, quantum bits for local metrology should be as small as possible and sensitive to specific fields of interest. While tremendous advances in QIS have been achieved in recent years, the underlying properties of the materials that host these quantum bits remains one of the key limitations in their performance.
Emerging scientific and engineering simulations running at scale on leadership-class High End Computing (HEC) environments are producing large volumes of data, which has to be transported and analyzed before any insights can result from these simulations. The complexity and cost (in terms of time and energy) associated with managing and analyzing this data have become significant challenges, and are limiting the impact of these simulations. Recently, data-staging approaches along with in-situ and in-transit analytics have been proposed to address these challenges by offloading I/O and/or moving data processing closer to the data. However, scientists continue to be overwhelmed by the large data volumes and data rates. In this paper we address this latter challenge. Specifically, we propose a highly scalable and low-overhead associative messaging framework that runs on the data staging resources within the HEC platform, and builds on the staging-based online in-situ/in- transit analytics to provide publish/subscribe/notification-type messaging patterns to the scientist. Rather than having to ingest and inspect the data volumes, this messaging system allows scientists to (1) dynamically subscribe to data events of interest, e.g., simple data values or a complex function or simple reduction (max()/min()/avg()) of the data values in a certain region of the application domain is greater/less than a threshold value, or certain spatial/temporal data features or data patterns are detected; (2) define customized in-situ/in-transit actions that are triggered based on the events, such as data visualization or transformation; and (3) get notified when these events occur. The key contribution of this paper is a design and implementation that can support such a messaging abstraction at scale on high- end computing (HEC) systems with minimal overheads. We have implemented and deployed the messaging system on the Jaguar Cray XK6 machines at Oak Ridge National Laboratory and the Lonestar system at the Texas Advanced Computing Center (TACC), and we present the experimental performance evaluation using these HEC platforms in the paper.