Bulletin of the American Physical Society
APS March Meeting 2019
Volume 64, Number 2
Monday–Friday, March 4–8, 2019; Boston, Massachusetts
Session A22: Building the Bridge to Exascale: Applications and Opportunities for Materials, Chemistry, and Biology IFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP DMP DCMP DCP Chair: Nichols Romero, Argonne Natl Lab Room: BCEC 157C |
Monday, March 4, 2019 8:00AM - 8:12AM |
A22.00001: Performance of Open-Source Real-Space Multigrid DFT Code on Pre-Exascale Architecture Emil Briggs, Wenchang Lu, Jerry Bernholc RMG (www.rmgdft.org) is an open-source suite of codes for performing large-scale, high-throughput electronic structure calculations. Designed for scalability, it discretizes the DFT equations on real-space grids that are distributed over the nodes of a massively parallel system via domain decomposition. The Kohn-Sham and Poisson equations are solved using multigrid techniques that dramatically accelerate convergence while only requiring nearest neighbor communications. In addition to the multigrid algorithms, the main parts of the calculations consist of dense matrix multiplications and iterative solutions of a partitioned eigenvalue problem that are particularly well suited for GPU accelerators. RMG makes very efficient use of GPUs, including multiple GPUs per node, if they are available. On the IBM/NVIDIA Summit supercomputer at ORNL, RMG utilizes all 6 GPUs per node and Cuda-managed memory to reach 83x performance improvement over the previous generation Cray XK7 Titan supercomputer at ORNL, which contains 1 prior-generation GPU per node. We will also discuss algorithmic improvements enabled by large-memory, high memory bandwidth nodes in view of future exascale architectures. |
Monday, March 4, 2019 8:12AM - 8:24AM |
A22.00002: Full-potential LSMS method for ab initio electronic structure calculations at large scale Yang Wang, Xianglin Liu, Markus Eisenbach, George Malcolm Stocks The locally self-consistent multiple scattering (LSMS) method is a linear scaling ab initio electronic structure calculation method in the framework of density functional theory with local density approximation. It is based on multiple scattering theory, which allows to use Green function and contour integration techniques for the calculation of electron density and density of states. With muffin-tin approximation, the LSMS method has demonstrated linear scaling and petascale performance at the scale of tens of thousands of atoms and has been applied to the study of nanostructures and random alloys. Recently, we have implemented full-potential capability in the LSMS method that enables the calculation of the Hellmann-Feynman force and also allows ab initio investigation of materials with dislocations, interstitial defects, etc. In this presentation, we will show the scalability of the full-potential LSMS method on supercomputers, and discuss its applications in the study of mechanical properties of transition metal alloys. |
Monday, March 4, 2019 8:24AM - 8:36AM |
A22.00003: Application of full-potential LSMS method in high entropy alloys Xianglin Liu, Yang Wang, Markus Eisenbach, George Malcolm Stocks The ability to treat large number of atoms from first principles is highly desired for the investigation of alloy systems. Here we demonstrate our latest implementation of the full potential LSMS method (FP-LSMS) that can accurately calculate systems with thousands of atoms. The acceleration is achieved by a parallelization of the FFT part of the Poisson solver for the electrostatic potential. We employed the code to investigate the elastic properties of high entropy alloys, which are a class of interesting new structural materials. Moreover, to demonstrate the power of the FP scheme, we used FP-LSMS to calculate the Hellman-Feynman forces. |
Monday, March 4, 2019 8:36AM - 8:48AM |
A22.00004: Opportunities for exascale in the characterization of uranium materials Ashley Shields, Andrew Miskowiec, Jennifer Niedziela, Marie C Kirkegaard, Roger Kapsimalis, Brian Anderson Characterizing new materials or unknown polycrystalline solids is a long-standing challenge for chemists. This problem has new resonance with the nuclear forensics community, where polycrystalline and amorphous materials are regularly encountered. Further, prediction of crystal structures and the calculation of vibrational properties for actinide materials is computationally intensive, limiting the unit cell size and number of structures which can be calculated. By analyzing basic geometric configurations, we hope to isolate spectroscopic characterization of low-symmetry systems to fundamental units. We have performed an analysis of coordination geometries of over 4000 uranium oxide and uranium fluoride structures, generated by combining genetic algorithm searches for stable phases with density functional theory. Here we present progress toward correlating vibrational properties with the fundamental coordination environments, aiming to facilitate interpretation of experimental analytical data from spectroscopy to characterize complex uranium phases. We discuss the opportunities and challenges for exascale computing and actinide materials, with an emphasis on the impacts and importance to the field of nuclear forensics. |
Monday, March 4, 2019 8:48AM - 9:00AM |
A22.00005: Moving the Effective Fragment Molecular Orbital method towards exascale Anastasia Gunina, Mark S Gordon Steep computational scaling and significant memory requirements of conventional electronic structure methods both limit the size of systems that can be described and hinder transitioning towards exascale computing. One of the ways to overcome these problems is to use fragmentation approaches. Our approach of choice is the Effective Fragment Molecular Orbitals (EFMO) method, which combines advantages of the Fragment Molecular Orbital (FMO) and the Effective Fragment Potential (EFP) methods. The current implementation in the GAMESS program package, however, is not optimized nor fully parallelized. |
Monday, March 4, 2019 9:00AM - 9:12AM |
A22.00006: Recent Advances in PARSEC for Performing Large-scale Electronic Structure Calculations in Real Space Kai-Hsin Liou, James Chelikowsky, Chao Yang The electronic structure of various types of materials, including complex biomolecules, nanostructures, and interfacial systems can be obtained by solving the Kohn–Sham (KS) equations. PARSEC, a real-space pseudopotential density functional theory KS equation solver, can routinely tackle systems with thousands of atoms using a Chebyshev-filtered subspace iteration (CheFSI) method. We will present a number of recent advances made in PARSEC that allow users to tackle larger systems that contain tens of thousands of atoms. These new advances include new parallelization strategies that significantly improve the scalability of orthogonalization and the Rayleigh–Ritz steps in the CheFSI framework and a more efficient way to perform the previously proposed spectrum slicing method. We will demonstrate the scalability of both the CheFSI and spectrum slicing methods on modern high performance computers and discuss the remaining challenges. |
Monday, March 4, 2019 9:12AM - 9:48AM |
A22.00007: Towards Paradigm Shifts in Electronic Structure Calculation for Large Systems: Wavelets, Fragments and Advanced Treatments of Excited States Invited Speaker: Laura Ratcliff The increasing power of massively parallel machines offers new opportunities for first principles materials simulations, providing software can be developed to effectively exploit new hardware. Density functional theory (DFT) has enjoyed widespread success for systems of up to a few hundred atoms, but is limited by the cubic scaling with the number of atoms of standard approaches. However, in recent years various linear scaling (LS) approaches have been developed, enabling simulations on tens of thousands of atoms. Since the parallel scalability is related to the number of atoms, such methods are also well suited to exploit supercomputers. One key factor influencing the accuracy and cost of DFT is the choice of basis set, where minimal, localized basis sets compete with extended, systematic basis sets. However, wavelets offer both locality and systematicity and are thus ideal for representing an adaptive local orbital basis which may be exploited for LS-DFT [1]. One may also make further approximations, e.g. dividing a system into fragments or exploiting underlying repetition of local chemical environments [2,3], where each approximation may be controlled and quantified. This ability to treat large systems with controlled precision offers the possibility of new types of materials simulations [4]. We will demonstrate the advantages of wavelets as a basis for large scale DFT calculations, as implemented in BigDFT. We will focus on the example of materials for organic LEDs, showing how our approach may be used to account for environmental and statistical effects on excited state calculations of disordered supramolecular materials [5]. |
Monday, March 4, 2019 9:48AM - 10:00AM |
A22.00008: Affordable and accurate large-scale hybrid-functional calculations on GPU-accelerated supercomputers Thierry DEUTSCH, Luigi Genovese, Laura Ratcliff Since 2008, the BigDFT project consortium has developed an ab initio DFT code based on Daubechies wavelets. |
Monday, March 4, 2019 10:00AM - 10:12AM |
A22.00009: Attacking the Strong and Weak Scaling Limits in Linear Scaling Hybrid Density Functional Theory Robert Distasio, Hsin-Yu Ko, Junteng Jia, Biswajit Santra, Zachary Sparrow, Alvaro Vazquez-Mayagoitia, Xifan Wu, Roberto Car Hybrid density functional theory (DFT) represents a quite favorable balance between accuracy and cost, and has therefore become the de facto standard in quantum chemistry. However, the steep computational cost associated with a cubic-scaling reciprocal-space evaluation of the exact-exchange energy has largely hindered the widespread use of hybrid DFT in the condensed phase. By utilizing a local representation of the occupied space, we have developed a formally exact and linear-scaling algorithm that exploits the real-space sparsity in the exact-exchange interaction by rigorously treating all overlapping orbital pairs. In this work, we present a series of theoretical developments and algorithmic improvements which drastically reduce the time to solution and take us one step closer to routine hybrid DFT-based ab initio molecular dynamics simulations of large-scale condensed-phase systems. In particular, we will focus on: (i) novel preconditioning techniques to speed up the computation, (ii) game and graph theoretical methods to improve the workload imbalance, (iii) sparse and asynchronous data transfer to mitigate the communication cost, and (iv) extensive code optimization/vectorization to efficiently utilize current- and next-generation architectures. |
Monday, March 4, 2019 10:12AM - 10:24AM |
A22.00010: Large-Scale Benchmark of Electronic Structure Solvers with the ELSI Infrastructure Victor Yu, William Dawson, Alberto Garcia, Ville Havu, Ben Hourahine, William P Huhn, Mathias Jacquelin, Weile Jia, Murat Keceli, Raul Laasner, Yingzhou Li, Lin Lin, Jianfeng Lu, Jose Roman, Alvaro Vazquez-Mayagoitia, Chao Yang, Volker Blum Routine application of electronic structure theory to systems consisting of thousands of atoms is often hindered by the solution of an eigenproblem. We here present an update to the ELectronic Structure Infrastructure (ELSI), an open-source software interface to facilitate the implementation and optimal use of high-performance solver libraries covering cubic scaling eigensolvers, linear scaling density-matrix-based algorithms, and other reduced scaling methods in between. The ELSI interface has been integrated into four electronic structure code projects (DFTB+, DGDFT, FHI-aims, SIESTA), forming the foundation of our effort to rigorously benchmark the performance of the solvers on equal footing. This presentation will particularly focus on a systematic set of large-scale benchmarks for multiple solvers performed with Kohn-Sham density-functional theory and density-functional tight-binding theory. Factors that strongly affect the efficiency of the solvers are identified and analyzed, including system size and dimensionality, matrix sparsity, eigenspectrum width, number of MPI processes, etc. Based on these benchmarks, we discuss our strategy to automatically select a solver for an arbitrary problem. |
Monday, March 4, 2019 10:24AM - 10:36AM |
A22.00011: Enabling Large-Scale Isobaric-Isothermal Hybrid Density Functional Theory Simulations in the Condensed Phase Hsin-Yu Ko, Robert Distasio, Biswajit Santra, Roberto Car The combination of ab initio molecular dynamics (AIMD) and high-performance computing (HPC) has the potential to furnish an atomistic-level understanding of complex condensed-phase systems such as molecular liquids and crystals. Such a detailed understanding requires an accurate treatment of both the quantum mechanical interactions and statistical mechanical sampling under realistic experimental conditions (e.g., in the isobaric-isothermal (NpT) ensemble). However, the routine use of sophisticated quantum mechanical methods like dispersion-inclusive hybrid density functional theory (DFT) is hindered by the cubic-scaling cost of conventional reciprocal-space based approaches. With the use of localized occupied orbitals, we have developed a formally exact and linear-scaling algorithm that directly addresses this prohibitive cost. In this work, we derive and implement the exact-exchange contributions to the stress tensor, thereby enabling hybrid DFT-based AIMD simulations of arbitrary cell sizes and shapes in the NpT ensemble. As an application of this method, we will discuss the pyridine molecular crystal. |
Monday, March 4, 2019 10:36AM - 10:48AM |
A22.00012: NWChemEx: Tackling Chemical, Materials, and Biochemical Challenges in the Exascale Era Wibe A De Jong, Kristopher W. Keipert, Raymond A. Bair, Ryan Richard, Jeffrey S. Boschen, Theresa L. Windus, Robert J Harrison, Thomas Dunning Historically, the NWChem computational chemistry package has been adapted and optimized for supercomputer architectures as they emerged. While this strategy has been successful over the past two decades, adapting NWChem to the major changes in hardware and programming models anticipated at the exascale is infeasible without a revolutionary redesign of the code. The NWChemEx project is redesigning and re-implementing NWChem for exascale computers using C++ with the goal of dramatically improving upon the scalability, performance, extensibility, and portability of the original application. In the first portion of this talk, the limitations of NWChem will be described. Next, we will illustrate how the scalability limitations are addressed in the NWChemEx design via the implementation of novel reduced-scaling methods with a new high-performance tensor library. Finally, we will show how lossy compression and differenced checkpointing techniques are utilized in NWChemEx to address more general exascale challenges related to data volume and system resilience. |
Monday, March 4, 2019 10:48AM - 11:00AM |
A22.00013: Fast Heat-Bath Configuration Interaction Junhao Li, Matthew Otten, Adam Holmes, Sandeep Sharma, Cyrus Jehangir Umrigar For many quantum mechanical systems the entire Hilbert space is enormous, but the important part is many orders of magnitude smaller. We have recently developed the semistochastic heat-bath configuration interaction (SHCI) method for efficiently calculating such systems. It is a systematically improvable selected configuration interaction plus perturbation theory method that is capable of giving essentially exact energies for larger systems than is possible with other such methods. I present recent advances we have made to the method that allow us to use 2 billion variational determinants and trillions of perturbative determinants. Details about the key data structures and the parallelization are presented. We use the algorithm to compute the potential energy surface of the very challenging chromium dimer in the cc-pVDZ-DK basis, correlating 28 valence and the semicore electrons. The Hilbert space has 5*1029 determinants. At equilibrium our energy agrees with the recent p-DMRG energy, but is more precise. We also present results for the homogeneous electron gas. |
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