Bulletin of the American Physical Society
APS March Meeting 2017
Volume 62, Number 4
Monday–Friday, March 13–17, 2017; New Orleans, Louisiana
Session A7: Computational Physics at the Petascale and Beyond IFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP DMP-DCMP DBIO-DCP Chair: Jack Wells, Oak Ridge National Laboratory Room: 266 |
Monday, March 13, 2017 8:00AM - 8:12AM |
A7.00001: ALCF Data Science Program: Productive Data-centric Supercomputing Nichols Romero, Venkatram Vishwanath The ALCF Data Science Program (ADSP) is targeted at “big data” science problems that require leadership computing resources. The goal of the program is to explore and improve a variety of computational methods that will enable data-driven discoveries across all scientific disciplines. The projects will focus on data science techniques covering a wide area of discovery including but not limited to uncertainty quantification, statistics, machine learning, deep learning, databases, pattern recognition, image processing, graph analytics, data mining, real-time data analysis, and complex and interactive workflows. Project teams will be among the first to access Theta, ALCF’s forthcoming 8.5 petaflops Intel/Cray system. The program will transition to the 200 petaflop/s Aurora supercomputing system when it becomes available. In 2016, four projects have been selected to kick off the ADSP. The selected projects span experimental and computational sciences and range from modeling the brain to discovering new materials for solar-powered windows to simulating collision events at the Large Hadron Collider (LHC). The program will have a regular call for proposals with the next call expected in Spring 2017.\linebreak http://www.alcf.anl.gov/alcf-data-science-program [Preview Abstract] |
Monday, March 13, 2017 8:12AM - 8:24AM |
A7.00002: BEAM: An HPC Pipeline for Nanoscale Materials Analysis and Neutron Data Modeling Eric Lingerfelt, Eirik Endeve, Yawei Hui, Chris Smith, Suhas Somnath, Nathan Grodowitz, Jose Borreguero, Feng Bao, Jennifer Niedziela, Dipanshu Bansal, Olivier Delaire, Richard Archibald, Alex Belianinov, Mallikarjun Shankar, Stephen Jesse The Bellerophon Environment for Analysis of Materials (BEAM) enables scientists at ORNL's Center for Nanophase Materials Sciences and Spallation Neutron Source to leverage the computational power of ORNL's Compute And Data Environment for Science (CADES) and the Oak Ridge Leadership Computing Facility (OLCF) to perform near real-time scalable analysis and modeling. At the core of this computational workflow system is a web and data server located at CADES that enables multiple, concurrent users to securely upload and manage data, execute materials science analysis and modeling workflows, and interactively explore results. BEAM's data management capabilities utilize a petabyte-scale file system and enable users to easily manipulate remote directories and uploaded data. The framework facilitates user workflow needs by enabling integration of advanced data analysis algorithms and push-button execution of dynamically generated HPC workflows employing these algorithms. We will present features of the system which include HPC analysis of SPM and STEM data and automated workflows for the optimization of inelastic and quasi-elastic neutron scattering data. [Preview Abstract] |
Monday, March 13, 2017 8:24AM - 8:36AM |
A7.00003: Extending Strong Scaling of Quantum Monte Carlo to the Exascale Luke Shulenburger, Andrew Baczewski, Ye Luo, Nichols Romero, Paul Kent Quantum Monte Carlo is one of the most accurate and most computationally expensive methods for solving the electronic structure problem. In spite of its significant computational expense, its massively parallel nature is ideally suited to petascale computers which have enabled a wide range of applications to relatively large molecular and extended systems. Exascale capabilities have the potential to enable the application of QMC to significantly larger systems, capturing much of the complexity of real materials such as defects and impurities. However, both memory and computational demands will require significant changes to current algorithms to realize this possibility. This talk will detail both the causes of the problem and potential solutions. [Preview Abstract] |
Monday, March 13, 2017 8:36AM - 9:12AM |
A7.00004: Here and now: the intersection of computational science, quantum-mechanical simulations, and materials science. Invited Speaker: Nicola Marzari The last 30 years have seen the steady and exhilarating development of powerful quantum-simulation engines for extended systems, dedicated to the solution of the Kohn-Sham equations of density-functional theory, often augmented by density-functional perturbation theory, many-body perturbation theory, time-dependent density-functional theory, dynamical mean-field theory, and quantum Monte Carlo. Their implementation on massively parallel architectures, now leveraging also GPUs and accelerators, has started a massive effort in the prediction from first principles of many or of complex materials properties, leading the way to the exascale through the combination of HPC (high-performance computing) and HTC (high-throughput computing). Challenges and opportunities abound: complementing hardware and software investments and design; developing the materials' informatics infrastructure needed to encode knowledge into complex protocols and workflows of calculations; managing and curating data; resisting the complacency that we have already reached the predictive accuracy needed for materials design, or a robust level of verification of the different quantum engines. In this talk I will provide an overview of these challenges, with the ultimate prize being the computational understanding, prediction, and design of properties and performance for novel or complex materials and devices. [Preview Abstract] |
Monday, March 13, 2017 9:12AM - 9:24AM |
A7.00005: Parameters Free Computational Characterization of Defects in Transition Metal Oxides with Diffusion Quantum Monte Carlo Juan A. Santana, Jaron T. Krogel, Paul R. Kent, Fernando Reboredo Materials based on transition metal oxides (TMO's) are among the most challenging systems for computational characterization. Reliable and practical computations are possible by directly solving the many-body problem for TMO's with quantum Monte Carlo (QMC) methods. These methods are very computationally intensive, but recent developments in algorithms and computational infrastructures have enabled their application to real materials. We will show our efforts on the application of the diffusion quantum Monte Carlo (DMC) method to study the formation of defects in binary and ternary TMO and heterostructures of TMO. We will also outline current limitations in hardware and algorithms. [Preview Abstract] |
Monday, March 13, 2017 9:24AM - 9:36AM |
A7.00006: High-efficiency wavefunction updates for large scale Quantum Monte Carlo Paul Kent, Tyler McDaniel, Ying Wai Li, Ed D'Azevedo Within ab intio Quantum Monte Carlo (QMC) simulations, the leading numerical cost for large systems is the computation of the values of the Slater determinants in the trial wavefunctions. The evaluation of each Monte Carlo move requires finding the determinant of a dense matrix, which is traditionally iteratively evaluated using a rank-1 Sherman-Morrison updating scheme to avoid repeated explicit calculation of the inverse. For calculations with thousands of electrons, this operation dominates the execution profile. We propose a novel rank-$k$ delayed update scheme. This strategy enables probability evaluation for multiple successive Monte Carlo moves, with application of accepted moves to the matrices delayed until after a predetermined number of moves, $k$. Accepted events grouped in this manner are then applied to the matrices en bloc with enhanced arithmetic intensity and computational efficiency. This procedure does not change the underlying Monte Carlo sampling or the sampling efficiency. For large systems and algorithms such as diffusion Monte Carlo where the acceptance ratio is high, order of magnitude speedups can be obtained on both multi-core CPU and on GPUs, making this algorithm highly advantageous for current petascale and future exascale computations. [Preview Abstract] |
Monday, March 13, 2017 9:36AM - 9:48AM |
A7.00007: OWL: A scalable Monte Carlo simulation suite for finite-temperature study of materials Ying Wai Li, Simuck F. Yuk, Valentino R. Cooper, Markus Eisenbach, Khorgolkhuu Odbadrakh The OWL suite is a simulation package for performing large-scale Monte Carlo simulations. Its object-oriented, modular design enables it to interface with various external packages for energy evaluations. It is therefore applicable to study the finite-temperature properties for a wide range of systems: from simple classical spin models to materials where the energy is evaluated by ab initio methods. This scheme not only allows for the study of thermodynamic properties based on first-principles statistical mechanics, it also provides a means for massive, multi-level parallelism to fully exploit the capacity of modern heterogeneous computer architectures. We will demonstrate how improved strong and weak scaling is achieved by employing novel, parallel and scalable Monte Carlo algorithms, as well as the applications of OWL to a few selected frontier materials research problems. [Preview Abstract] |
Monday, March 13, 2017 9:48AM - 10:00AM |
A7.00008: Real-time electron dynamics for massively parallel excited-state simulations Xavier Andrade The simulation of the real-time dynamics of electrons, based on time dependent density functional theory (TDDFT), is a powerful approach to study electronic excited states in molecular and crystalline systems. What makes the method attractive is its flexibility to simulate different kinds of phenomena beyond the linear-response regime, including strongly-perturbed electronic systems and non-adiabatic electron-ion dynamics. Electron-dynamics simulations are also attractive from a computational point of view. They can run efficiently on massively parallel architectures due to the low communication requirements. Our implementations of electron dynamics, based on the codes Octopus (real-space) and Qball (plane-waves), allow us to simulate systems composed of thousands of atoms and to obtain good parallel scaling up to 1.6 million processor cores. Due to the versatility of real-time electron dynamics and its parallel performance, we expect it to become the method of choice to apply the capabilities of exascale supercomputers for the simulation of electronic excited states. [Preview Abstract] |
Monday, March 13, 2017 10:00AM - 10:12AM |
A7.00009: Massively Parallel Real-Time TDDFT Simulations of Electronic Stopping Processes Dillon Yost, Cheng-Wei Lee, Erik Draeger, Alfredo Correa, Andre Schleife, Yosuke Kanai Electronic stopping describes transfer of kinetic energy from fast-moving charged particles to electrons, producing massive electronic excitations in condensed matter. Understanding this phenomenon for ion irradiation has implications in modern technologies, ranging from nuclear reactors, to semiconductor devices for aerospace missions, to proton-based cancer therapy. Recent advances in high-performance computing allow us to achieve an accurate parameter-free description of these phenomena through numerical simulations. Here we discuss results from our recently-developed large-scale real-time TDDFT implementation for electronic stopping processes in important example materials such as metals, semiconductors, liquid water, and DNA. We will illustrate important insight into the physics underlying electronic stopping and we discuss current limitations of our approach both regarding physical and numerical approximations. This work is supported by the DOE through the INCITE awards and by the NSF. Part of this work was performed under the auspices of U.S. DOE by LLNL under Contract DE-AC52-07NA27344. [Preview Abstract] |
Monday, March 13, 2017 10:12AM - 10:24AM |
A7.00010: Toward Petascale Biologically Plausible Neural Networks Lyle Long This talk will describe an approach to achieving petascale neural networks. Artificial intelligence has been oversold for many decades. Computers in the beginning could only do about 16,000 operations per second. Computer processing power, however, has been doubling every two years thanks to Moore's law, and growing even faster due to massively parallel architectures. Finally, 60 years after the first AI conference we have computers on the order of the performance of the human brain (10$^{16}$ operations per second). The main issues now are algorithms, software, and learning. We have excellent models of neurons, such as the Hodgkin-Huxley model, but we do not know how the human neurons are wired together. With careful attention to efficient parallel computing, event-driven programming, table lookups, and memory minimization massive scale simulations can be performed. The code that will be described was written in C$++$ and uses the Message Passing Interface (MPI). It uses the full Hodgkin-Huxley neuron model, not a simplified model. It also allows arbitrary network structures (deep, recurrent, convolutional, all-to-all, etc.). The code is scalable, and has, so far, been tested on up to 2,048 processor cores using 10$^{7}$ neurons and 10$^{9}$ synapses. [Preview Abstract] |
Monday, March 13, 2017 10:24AM - 10:36AM |
A7.00011: Machine learning properties of materials and molecules with entropy-regularized kernels Michele Ceriotti, Albert Bartók, Gábor Csányi, Sandip De Application of machine-learning methods to physics, chemistry and materials science is gaining traction as a strategy to obtain accurate predictions of the properties of matter at a fraction of the typical cost of quantum mechanical electronic structure calculations. In this endeavor, one can leverage general-purpose frameworks for supervised-learning. It is however very important that the input data -- for instance the positions of atoms in a molecule or solid -- is processed into a form that reflects all the underlying physical symmetries of the problem, and that possesses the regularity properties that are required by machine-learning algorithms. Here we introduce a general strategy to build a representation of this kind. We will start from existing approaches to compare local environments (basically, groups of atoms), and combine them using techniques borrowed from optimal transport theory, discussing the relation between this idea and additive energy decompositions. We will present a few examples demonstrating the potential of this approach as a tool to predict molecular and materials' properties with an accuracy on par with state-of-the-art electronic structure methods. [Preview Abstract] |
Monday, March 13, 2017 10:36AM - 10:48AM |
A7.00012: Improving the network efficiency of the two particle parquet algorithm Samuel Kellar, Bibek Wagle, Ka-Ming Tam, Hartmut Kaiser, Juana Moreno, Mark Jarrell Strongly correlated systems require large scale simulations. Perturbative methods such as the two particle self-consistent parquet algorithm require the storage of large rank-three vertex functions. This data is transferred via an all-to-all communication which is costly since network transfer of data is orders of magnitude slower than floating point operations. The size of the messages compounds the issue by introducing large overheads. An analysis of the vertices reveals large amounts of noise. This enables significant compression and opportunities for message coalescing which reduces the network traffic resulting in a significant speedup. These ideas should be broadly applicable to other problems which involve large scale data transfer. [Preview Abstract] |
Monday, March 13, 2017 10:48AM - 11:00AM |
A7.00013: Computational studies of the 2D self-assembly of bacterial microcompartment shell proteins Jyoti Mahalik, Kirsten Brown, Xiaolin Cheng, Miguel Fuentes-Cabrera Bacterial microcomartments (BMCs) are subcellular organelles that exist within wide variety of bacteria and function like nano-reactors. Among the different types of BMCs known, the carboxysome has been studied the most. The carboxysomes plays an important role in the transport of metabolites across its outer proteinaceous shell. Plenty of studies have investigated the structure of this shell, yet little is known about its self-assembly . Understanding the self-assembly process of BMCs' shell might allow disrupting their functioning and designing new synthetic nano-reactors. We have investigated the self-assembly process of a major protein component of the carboxysome's shell using a Monte Carlo technique that employed a coarse-grained protein model that was calibrated with the all-atomistic potential of mean force. The simulations reveal that this protein self-assembles into clusters that resemble what were seen experimentally in 2D layers. Further analysis of the simulation results suggests that the 2D self-assembly of carboxysome's facets is driven by nucleation-growth process, which in turn could play an important role in the hierarchical self-assembly of BMCs' shell in general. [Preview Abstract] |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700