Colloquia & Seminars
PICSciE Seminar: Exascale and Extreme Data Science at NERSC
Sudip Dosanjh, NERSC Director
Tuesday, April 18, 4:00 – 5:00 PM
Vis Lab, 347 Lewis Science Library
Washington Road & Ivy Lane
The National Energy Research Scientific Computing Center’s primary mission is to accelerate scientific discovery at the U.S. Department of Energy's Office of Science through high performance computing and data analysis. NERSC supports the largest and most diverse research community of any computing facility within the DOE complex, providing large-scale, state-of-the-art computing for DOE’S unclassified research programs in alternative energy sources, environmental science, materials research, astrophysics and other science areas related to DOE’s science mission.
Cori, NERSC’s new supercomputer, is deployed in Berkeley Laboratory’s new Computational Research and Theory (CRT) Facility. It has over 9300 manycore Intel Knight’s Landing processors, which introduce several technological advances, including higher intra-node parallelism; high-bandwidth, on-package memory; and longer hardware vector lengths. These enhanced features are expected to yield significant performance improvements for applications running on Cori. In order to take advantage of the new features, however, application developers will need to make code modifications because many of today’s applications are not optimized to take advantage of the manycore architecture and on-package memory.
Cori includes many enhancements to enable a rapidly growing extreme data science workload at NERSC. Cori has a 2000 Intel® Haswell processor partition with larger memory nodes to enable extreme data analysis. A fast internet connection lets users stream data from experimental and observational facilities directly into the system. A “Burst Buffer”, a 1.5 Petabyte layer of NVRAM, helps accelerate I/O. Cori also includes a number of software enhancements to enable complex workflows. For the longer term we are investigating whether a single system can meet the simulation and data analysis requirements of our users.
Dr. Sudip Dosanjh is Director of NERSC at Lawrence Berkeley National Laboratory. Previously, Dr. Dosanjh headed extreme-scale computing at Sandia National Laboratories. He was co-director of the Los Alamos/Sandia Alliance for Computing at the Extreme-Scale from 2008-2012. He also served on the U.S. Department of Energy’s Exascale Initiative Steering Committee for several years. He had a key role in establishing co-design as a methodology for reaching exascale computing. He has numerous publications on exascale computing, co-design, computer architectures, massively parallel computing and computational science.
Machine learning based intelligent earthquake data processing for global adjoint tomography
Yangkeng Chen, Research Associate, Oak Ridge National Laboratory.
Monday, April 17, 12:00 – 1:00 PM
Vis Lab, 347 Lewis Science Library
Washington Road & Ivy Lane
Due to the increased computational capability afforded by modern and future computing architectures, the seismology community is demanding a more comprehensive understanding of the full waveform information from the recorded earthquake seismograms. Global adjoint tomography is a complex workflow that matches observed seismic data with synthesized seismograms by iteratively updating the earth model parameters based on the adjoint state method. This methodology allows us to compute a very accurate model of the earth's interior. The synthetic data is simulated by solving the wave equation in the entire globe using a spectral-element method. In order to ensure the inversion accuracy and stability, both the synthesized and observed seismograms must be carefully pre-processed. Because the scale of the inversion problem is extremely large and there is a very large volume of data to both be read and written, an efficient and reliable pre-processing workflow must be developed. We are investigating intelligent algorithms based on a machine learning framework that will automatically tune parameters for the data processing chain. In the current machine learning framework, optimal misfit calculation windows in the seismograms can be automatically detected and thus extremely noisy and deviated waveforms are deserted, just like the face recognition in many computer vision applications. The intelligent earthquake data processing framework will enable the seismology community to compute the global adjoint tomography using seismic data from an arbitrarily large number of earthquake events in the fastest, most efficient way.
Yangkang Chen received the B.S. degree in geophysics from the China University of Petroleum, Beijing, in 2012, and the Ph.D. degree in geophysics from the University of Texas at Austin in 2015. He is currently a Distinguished Post-Doctoral Research Associate with the Oak Ridge National Laboratory. His long-time research interest includes machine learning, large-scale seismic data processing and inversion. His PhD thesis focused on high-resolution seismic imaging for oil&gas exploration and reservoir monitoring. His is now devoted to intelligently harnessing massive earthquake data for obtaining an unprecedented high-resolution global earth picture. Dr. Chen is a high-impact scholar with a strong publication record and serves many internationally renowned conferences and journals as an editor, a chair, and a reviewer.
Lunch will be provided.
High Performance Computing Paradigms in Python
Julian Kates-Harbeck, Harvard University
Thursday, April 6, 12:00 – 1:00 pm
138 Lewis Science Library
(Lunch will be provided at 11:45am outside of the lecture hall)
In this talk, we will give an overview of several key paradigms for high performance computing in python: GPU computing, MPI, and multiprocessing. Relevant application areas include scientific computing and machine learning at scale. Other topics we will cover include key python packages and tradeoffs for when to use a given approach. We will show real world code examples that make use of these approaches, and work through an interactive demonstration of some simple examples of speeding up serial code.
Julian Kates-Harbeck received his bachelor’s degree in physics at Stanford and later earned a master’s in computer science (with a focus on AI and machine learning) at the same university. In the period in between, he co-founded a tech startup, where he was responsible for hiring, product and strategy. He is currently pursuing a PhD in biophysics (specifically evolutionary dynamics on networks) at Harvard. For his studies there, he was awarded the National Science Foundation GRFP, Department of Defense NDSEG, and Department of Energy CSGF fellowships. Julian has written about biophysics, machine learning, astrophysics and plasma physics.
PICSciE Colloquium: The U.S. D.O.E. Exascale Computing Project – Goals and Challenges
Exascale Computing Project Director
Argonne Distinguished Fellow
Argonne National Laboratory
Wednesday, March 29th, 11:00 am – 12:00 pm
120 Lewis Science Library, Washington Road & Ivy Lane
The U.S. Department of Energy established in 2016 the Exascale Computing Project (ECP) -- a joint project of the DOE Office of Science (DOE-SC) and the DOE National Nuclear Security Administration (NNSA) -- that will result in a capable exascale ecosystem and prepare mission critical scientific and engineering applications to take advantage of that ecosystem.
This presentation will describe the goals of the ECP, its plans for achieving them, the challenges to be overcome, and its current status, as well as what elements of the exascale ecosystem are outside of the scope of the ECP.
Refreshments will be provided.
Disruption Forecasting in Tokamak Fusion Plasmas using Deep Recurrent Neural Networks
Lunch will be served from 11:45-12:00 pm ∙
The prediction and avoidance of disruptions in tokamak fusion plasmas represents a key challenge on the way to stable energy production from nuclear fusion. A fusion plasma is a complex dynamical system with some unknown internal state which emits a time series of possibly high dimensional observable data that is captured by sensory diagnostics. Using such diagnostic data from past plasma shots with both disruptive and non-disruptive outcomes, we train a deep recurrent neural network to predict the onset of disruptions in an online setting. To deal with very large amounts of data and the need for iterative hyperparameter tuning, we also introduce a distributed training algorithm that runs on MPI clusters of GPU nodes and provides strong linear runtime scaling. Our approach demonstrates competitive predictive performance on experimental data from the JET tokamak, and we highlight promising avenues for extending our method to cross-tokamak prediction as well as to high-dimensional diagnostic data such as temperature and density profiles.
Julian grew up in Munich, Germany. He got his bachelor’s degree in physics at Stanford and later earned a master’s in computer science (with a focus on AI and machine learning) at the same university. In the period in between, he co-founded a tech startup, where he was responsible for hiring, product and strategy. He is currently pursuing a PhD in biophysics (specifically evolutionary dynamics on networks) at Harvard. For his studies there, he was awarded the National Science Foundation GRFP, Department of Defense NDSEG, and Department of Energy CSGF fellowships. Julian has written about biophysics, machine learning, astrophysics and plasma physics.
Enabling Scale-Up, Scale-Out, and Scale-Deep for Big Data
Dr. Jeremy Kepner
MIT Lincoln Laboratory Fellow
Head, Lincoln Laboratory Supercomputing Center
Monday, October 10, 2016, 12:00 - 1:00 pm
120 Lewis Science Library, Washington Road and Ivy Lane
Next Generation Applications: Using a Productivity Focus, 9/23/15
Michael A. Heroux
Distinguished Member of Technical Staff, Sandia National Labs
Scientist in Residence, St John’s University
The extreme-scale computing community is several years into a highly disruptive period of change. New commodity performance curves must be incorporated into application designs, and the orders of magnitude in performance potential will increase the demand to couple physics and scales into a single integrated execution environment.
In this talk we discuss a productivity focus as the fundamental source for guiding application development activities. Although productivity is always implicitly part of our decisions, an explicit focus on it may lead to new activities, and strategies we have not seriously considered before. We will talk about emerging application architectures, development and use of software ecosystems, software best practices and characterize some of the important attributes of future scalable applications.
Michael Heroux is a Distinguished Member of the Technical Staff at Sandia National Laboratories and Scientist in Residence at St. John’s University, MN, working on new algorithm development, and robust parallel implementation of solver components for problems of interest to Sandia and the broader scientific and engineering community. He leads development of the Trilinos Project, an effort to provide state of the art solution methods in a state of the art software framework. Dr. Heroux works on the development of scalable parallel scientific and engineering applications and maintains his interest in the interaction of scientific/engineering applications and high performance computer architectures. He leads the Mantevo project, which is focused on the development of Open Source, portable mini-applications and mini-drivers for scientific and engineering applications. Dr. Heroux is also the lead developer and architect of the HPCG benchmark, intended as an alternative ranking for the TOP 500 computer systems.
Dr. Heroux is a member of the Society for Industrial and Applied Mathematics (SIAM) and past chair of the SIAM Activity Group on Supercomputing. He is a Distinguished Member of the Association for Computing Machinery (ACM). He is the Editor-in-Chief for the ACM Transactions on Mathematical Software, Subject Area Editor for the Journal on Parallel and Distributed Computing and Associate Editor for the SIAM Journal on Scientific Computing.
Software Challenges for Extreme Scale Systems
Professor Vivek Sarkar
What does Titan tell us about preparing for exascale supercomputers?
Speaker: Jack Wells, Director of Science, Oak Ridge National Labs
Monday, February 10, 2014
12-1:30pm, Vis Lab, Room 346 Lewis Library
Lunch will be provided at Noon
Modeling and simulation with petascale computing has supercharged the process of innovation, dramatically accelerating time-to-insight and time-to-discovery. The Titan supercomputer is the Department of Energy’s flagship Cray XK7 managed by the Oak Ridge Leadership Computing Facility (OLCF). With its hybrid, accelerated architecture of traditional CPUs and graphics processing units (GPUs), Titan allows advanced scientific applications to reach speeds exceeding 10 petaflops with a marginal increase in electrical power demand over the previous generation leadership-class supercomputer. I will summarize the lessons learned in deploying Titan and in preparing applications to move from conventional CPU architectures to a hybrid, accelerated architectures, with a focus on early science outcomes from Titan. We will discuss implications for the research community as we prepare for exascale computational science and engineering within the next decade. I will also provide an overview of user programs at the Oak Ridge Leadership Computing Facility with specific information how researchers may apply for allocations of computing resources.
Jack Wells is the Director of Science for the National Center for Computational Sciences (NCCS) at Oak Ridge National Laboratory (ORNL) with the rank of Distinguished R&D Scientist. He is responsible for devising the strategy to ensure cost-effective, state-of-the-art scientific computing at the NCCS, which hosts the Department of Energy’s Oak Ridge Leadership Computing Facility (OLCF), a national user facility, and Titan, currently the faster supercomputer in the United States. Dr. Wells began his ORNL career in 1990 for resident research on his Ph.D. in Physics from Vanderbilt University. Following a three-year postdoctoral fellowship at the Harvard-Smithsonian Center for Astrophysics, he returned to ORNL as a staff scientist in 1997 as a Wigner fellow. Jack is an accomplished practitioner of computational physics and has been sponsored in his research by the Department of Energy’s Office of Basic Energy Sciences.
MAE/PICSciE seminar: “Stability Analysis of an Impacting T-Junction Pipe Flow”
Mechanical and Aerospace Engineering
Monday, December 9, 2013
12:00 – 1:00 pm, EQuad J223
An asymptotic parallel-in-time method for highly oscillatory PDEs
Smagorinsky Room, NOAA/GFDL
Forrestal Campus, 201 Forrestal Road
Algorithmic requirements for extreme scale simulation
Speaker: David E. Keyes, Professor, Applied Mathematics and Computational Science, Director, Strategic Initiative in Extreme Computing, King Abdullah University of Science and Technology (KAUST)
Date: February 7, 2013 from 1:30 pm - 2:30 pm, Visualization Lab, 346 Lewis Science Library
Light refreshments will be served
Diverging exponentials in computer hardware subsystem performance require rethinking of models and reimplementation of algorithms in scientific and engineering simulation. Much mathematics and software appears to be missing if emerging hardware is to be used near its potential, since our existing code base has been assembled with a premium on squeezing out flops and improving the execution rate of those that remain. Instead, for reasons of energy efficiency and system acquisition cost, we must now focus on squeezing out synchronizations, memory footprint, and memory transfers. High concurrency and power-efficient design of the individual cores put opposite pressures on algorithms: respectively, they require greater data locality and greater freedom to redistribute data and computation. After decades of programming model stability, new models and new hardware must be developed simultaneously, a process called co-design. We extrapolate current trends and describe directions for exascale algorithms.
The Astronomical Multipurpose Software Environmentand the Ecology of Star Clusters
Speaker: Simon Portegies Zwart, Professor of Computational Astrophysics at the Sterrewacht Leiden of Leiden University
Date: February 13, 2012 from 12:30-1:30pm, Room Location TBD
Light Refreshments will be served at 11:45am in the PICSciE Reception area
Star cluster ecology is the field of research where stellar evolution, gravitational dynamics, hydrodynamcs and the background potential dynamics of the parent galaxy interact to a complex non-linear evolution of self gravitating stellar systems. I will review the processes related to the ecology of stellar clusters, discuss the numerical hurdles and the physical principles. In addition, I will introduce the AMUSE framework with which we are performing simulations of the ecology of stellar clusters. AMUSE is a general purpose framework for interconnecting existing scientific software with a homogeneous and unified interface. The framework is based on the standard message passing interface any production ready code that is written in a language that supports its native bindings can be incorporated, in addition our framework is intrinsically parallel and it conveniently separates the all the numerical solvers in memory.
Simon Portegies Zwart was born in Amsterdam and studied astronomy at the University of Amsterdam. After his PhD with Frank Verbunt at Utrecht University he traveled over the world while working as a postdoctoral fellow at the University of Amsterdam, Tokyo University (Japan), MIT (USA) and back to Amsterdam. He currently is full professor of computational astrophysics at the Sterrewacht Leiden of Leiden University. His professional interests are high-performance computing and gravitational stellar dynamics, in particular the ecology of dense stellar systems. His personal interests include translating Egyptian hieroglyphs and brewing beer.
Perspectives on China’s Role in Global High Performance Computing
Date: January 23, 2012 from 12:30-1:30pm, 121 Lewis Science Library
Light Refreshments will be served at 11:45 in the PICSciE Reception area.
High performance computing is generally recognized to be an increasingly vital tool for accelerating progress in scientific research in the 21st Century. China’s rapid emergence in this area has been remarkable, and the current presentation will highlight associated impressions from a number of visits there by me and national colleagues over the past year. At the top of the most recent LINPACK list (November, 2011) are Japan’s Fujitsu K machine at No. 1, the Chinese supercomputers at Nos. 2 and 4, and the U.S. falling to No. 3. It is significant to note that the 2.57 petaflops performance level of the Chinese Tianhe-1A system at the National Supercomputer Center in Tianjin has passed the U.S.’s Cray XT5 system “Jaguar” at the Oak Ridge National Laboratory with 1.76 petaflops – previously the No. 1 machine in June 2010. The rapid rise of HPC hardware in China over the past decade is particularly notable since the Chinese systems, which were basically absent from the Top500 list prior to 2001, now occupy the Nos. 2 and 4 positions.
William M. Tang is the Director of the Fusion Simulation Program at the Princeton Plasma Physics Laboratory (PPPL) and serves on the Executive Committee for PICSciE which he helped establish during his 6-years as Associate Director. He is a Fellow of the American Physical Society, and in October, 2005, received the Chinese Institute of Engineers-USA (CIE-USA) Distinguished Achievement Award “for his outstanding leadership in fusion research and contributions to fundamentals of plasma science. He was the Chief Scientist at PPPL from 1997 until 2009 and also played a national leadership role in the formulation and development of the DoE’s multi-disciplinary program in advanced scientific computing applications, SciDAC (Scientific Discovery through Advanced Computing). He chaired the major DoE-SC meeting on “Scientific Grand Challenges in Fusion Energy Sciences and the Role of Computing at the Extreme Scale” (Spring, 2009).
Campus-Scale High Performance Cyberinfrastructure for Data-Intensive Research
Blue Waters Plus – A Super System to Solve Super Challenges
Understanding the Human Brain: The Ultimate Computational Challenge (in Theory and Practice)
Speaker: Professor Jonathan Cohen, Princeton Neuroscience Institute
April 11, 2011 from 12:30 pm - 1:30 pm, Visualization Lab, 346 Lewis Library
The human brain is the most complex device in the known universe. With an estimated 100 billion neurons, 100 trillion connections among them, and an inestimable number of potential circuits, the challenge to track these and understand their function is arguably the greatest challenge science has ever faced. It is a trivial assertion, therefore, that this challenge demands the most sophisticated approaches to mathematical analysis and numerical (computational) simulation we can garner. This is true both for theory development, as well as for data analysis. The former stems from the inherent complexity of the problem, and the latter from the size of the datasets required to make progress in addressing it. I will review the state-of-the along these dimensions, focusing in particular on the challenge posed by analyzing human brain imaging data — the most available measures we have of the functioning of the intact human brain.
Computational approaches to the study of collective behavior
Speaker: Prof. Iain Couzin, Department of Ecology & Evolutionary Biology, Princeton University
March 28, 2011 from 12:30 pm - 1:30 pm, Visualization Lab, 346 Lewis Library
A fundamental problem in a wide range of biological disciplines is understanding how functional complexity at a macroscopic scale (such as the functioning of a biological tissue) results from the actions and interactions among the individual components (such as the cells forming the tissue). Animal groups such as bird flocks, fish schools and insect swarms frequently exhibit complex and coordinated collective behaviors and present unrivaled opportunities to link the behavior of individuals with the functioning and efficiency of dynamic group-level properties.
Using an integrated experimental and theoretical approach involving both insects and vertebrates I will address both how, and why, animals coordinate behavior., and the computational tools that we have developed to facilitate their study. In some animal groups decision-making by individuals is so integrated that it has been associated with the concept of a “collective mind”. Since each organism has relatively local sensing ability, coordinated animal groups have evolved collective strategies that allow individuals to access higher-order computational abilities at the group level. I investigate the coupling between spatial and information dynamics in swarms, flocks, schools and herds and reveal the critical role uninformed individuals (those who have no information about the feature upon which a collective decision is being made) play in inhibiting extremism and promoting democratic consensus in groups.
Toward Exascale Computing in Gyrokinetic Particle-in-Cell Simulations of Fusion Plasmas
Speaker: Stephane Ethier, Computational Scientist, PPPL
March 24, 2011 from 12:30 pm - 1:30 pm, Visualization Lab, 346 Lewis Library
The last decade has witnessed a rapid emergence of larger and faster computing systems in the US supercomputing centers. Massively parallel machines have gone mainstream and are now the tool of choice for large scientific simulations. Scientific applications need to be modified, adapted, and optimized for each new system being introduced. With a few petascale systems now in production mode, the focus of the DOE Office of Advanced Scientific Computing Research has now shifted to the next level of "Exascale", which promises to be truly disruptive. With an estimated billion cores to deal with, scientific applications will need to manage extreme parallelism, limited bandwidth, frequent failures, and many more hardware and software challenges. In this talk, I will discuss the path to extreme scale computing from the point of view of the large-scale gyrokinetic particle-in-cell codes developed at Princeton University's Plasma Physics Laboratory to study microturbulent transport in fusion plasmas
3D Visualization and Physically-based Illumination
Speaker: David Banks, University of Tennessee and Oak Ridge National Laboratory
David Banks holds positions as tenured faculty in the EECS department at the University of Tennessee and as senior scientist in scientific computing at Oak Ridge National Laboratory. He is a member of the UT/ORNL Joint Institute for Computational Sciences, home to the top-ranked academic supercomputer in the world.
February 28, 2011 from 12:30 pm - 1:30 pm, Visualization Lab, 346 Lewis Library
"3D data visualization" applies computer graphics to datasets of various kinds. Graphics algorithms can be viewed as "solvers" for radiation transport. Our lab investigates the interplay among transport, rendering, visualization, and human perception. We have found that perception of 3D scenes can be improved by visualizing them using rendering algorithms that more accurately solve the transport equation. Surprisingly, such "physically-based” algorithms have not been widely adopted by scientific users.
New frontiers in quantum chemistry using supercomputers
Speaker: Jeffrey Hammond, University of Chicago / Argonne National Laboratory
Jeff Hammond is currently a Director's Postdoctoral Fellow at the Argonne Leadership Computing Facility. He received his PhD in chemistry from the University of Chicago as a DOE Computational Science Graduate Fellow.
February 21, 2011 from 12:30 pm - 1:30 pm, Visualization Lab, 346 Lewis Library
Recent advancements in high-performance computing present both challenges and new opportunities for quantum chemists. Accurate methods like coupled-cluster theory can now be applied to systems with
dozens of atoms, opening up new application areas related to biology and material science. I will present recent results obtained using the massively parallel quantum chemistry package NWChem, highlighting
the importance of accurate many-body simulations of electric-field response properties and electronic excited-states for a diverse set of chemical systems. The rigorous development of force-fields, including both inter- and intramolecular terms, will also be discussed. Finally, I will discuss recent developments in computer architecture - million-way parallelism and heterogeneous nodes - affect algorithms and software development in correlated electronic structure calculations.
Extracting Biological Insight from Complex Genome-Scale Data: Connecting Growth Control and Stress Response in Yeast
Speaker: David Botstein, Director, Lewis-Sigler Institute for Integrative Genomics
February 14, 2011 from 12:30 pm - 1:30 pm, Visualization Lab, 346 Lewis Library
The maintenance of cellular homeostasis in the face of rapidly changing environmental conditions has been the focus of our research for the past five years. Specifically, we have studied the relationship between the growth rate, which we can control directly by setting the dilution rate in chemostats, and the initiation of cell division cycle, response to environmental stress, and metabolism. We have exploited high-throughput methods, some of our own devising, to follow gene expression, metabolite levels, and relative fitness of mutants on a comprehensive scale in order to obtain a view of the integration of these functions at the system level.
The biggest challenge in this kind of research is not the acquisition, nor even the statistical analysis of the data. Instead, it is the visualization of the analysis of the results in a form that can be appreciated and communicated by scientists. Examples will be provided from our research that illustrate this challenge and some of the ways we have attempted to meet it.