InTheLoop | 04.15.2013
GlobusWorld Conference Features Presentations by NERSC, ESnet Staff
The 11th annual GlobusWORLD conference, to be held April 16–18 at Argonne National Laboratory, will feature presentations by David Skinner and Shreyas Cholia of NERSC and Eli Dart of ESnet. The conference is built around Globus Online, an easy-to-use, highly reliable online file transfer and sharing service.
Globus Online has become a primary on-ramp for researchers to access high performance networks like ESnet for rapidly sharing data or to use remote computing facilities like NERSC. This year’s conference focuses on "moving, syncing and sharing" research data at scale.
Berkeley Lab speakers are:
David Skinner, leader of NERSC's Outreach, Software and Programming Group, will discuss "Web APIs for the Big Science Enterprise." According to Skinner, while the Internet has transformed how companies interact and leverage their resources, too much of the scientific enterprise remains hidden behind outdated interfaces. Building reliable pipes to connect scientific resources and disseminate data is often still done manually and lacks a programmatic interface. Science-focused web Application Programming Interfaces (APIs) provide a path forward that has broad potential benefits to science teams both large and small, their management stakeholders, and the public.
Eli Dart, an ESnet network engineer, will describe "Optimizing Data Management at the Advanced Light Source with a Science DMZ." ESnet is working with scientists at the Advanced Light Source at Berkeley Lab who are seeing massive increases in the data output of their experiments and who now require HPC and network resources to support their research. Dart's talk details ESnet's work with an ALS scientist who needed new resources to handle a 50-fold increase in data generation. By deploying a dedicated Science DMZ infrastructure and using Globus Online, the scientist now seamlessly transmits data to NERSC, where the data is stored, managed and shared with other researchers.
Shreyas Cholia, a computer systems engineer at NERSC and the Department of Energy Systems Biology Knowledgebase (KBase), will talk about "Using the Globus Online Platform at KBase." KBase is a software and data environment designed to enable researchers to collaboratively generate, test and share new hypotheses about gene and protein functions; perform large-scale analyses; and model interactions in microbes, plants and their communities. KBase and Globus Online have formed a collaboration that simplifies KBase’s account management by providing simple, reliable mechanisms for creating and managing user accounts and groups.
Sudip Dosanjh Discusses Path to Exascale at Edinburgh Conference
NERSC Director Sudip Dosanjh participated in a panel on "The Path to Exascale" at the Exascale Applications and Software Conference held April 9–12 in Edinburgh, Scotland. He discussed NERSC's plans for exascale and some of the technical challenges facing the high performance computing community. The conference was hosted by the Edinburgh Parallel Computing Center. Also participating on the panel were Jack Dongarra (University of Tennesse), Satoshi Matsuoka (Tokyo Institute for Technology), Barbara Chapman (University of Houston), and Peter Oster (Finnish IT Center for Science).
April 18 Workshop on Career and Leadership Strategies
It takes certain skills and strategies to maximize scientific career opportunities and develop as a leader — but according to Peter Fiske, none of them are taught in grad school. Attend Stepping Up: Practical Career and Leadership Advice for Scientists (BLI1072) on Thursday, April 18, from noon to 1:30 p.m. in the Building 50 auditorium, to hear Fiske's insightful perspectives. The author of "Put Your Science to WORK: A Take-Charge Career Guide for Scientists" and a career columnist for Science and Nature, Fiske regularly conducts workshops at UC Berkeley, Harvard, and Stanford. He is also a serial entrepreneur and CEO of PAX Water Technologies, Inc.
This course is designed for postdocs and GSRAs. Scientists are also welcome to attend. Register for this training here. Click on Training Enrollment, then Employee Skills Development, then select BLI1072, Science Careers & Leadership.
This Week’s Computing Sciences Seminars
Computation-Guided Experimentation in Materials Research: A Systems-Level Approach to Science for Innovation
Tuesday, April 16, 12:00–1:00 pm, 15-253 (User Support Building)
Julie Christodoulou, Director, Naval Materials Division, U.S. Office of Naval Research
Materials science and engineering has long been a focus of efforts funded by the Office of Naval Research, which has a unique responsibility of supporting basic and applied research. This introductory seminar will describe the materials program, emphasizing the role of integrated computation and experimentation to elucidate the fundamental understanding that empowers innovative solutions. The Dynamic 3D Digital Structure program, which aimed to change approaches to the way basic research is conducted, will be highlighted to provide good examples of how the terms “mission-driven” and “fundamental research” need not be mutually exclusive. How computation-guided experimentation is a key enabler to achieving a systems-level approach to science and engineering will be explored and reconciled with the goals of the Materials Genome Initiative for Global Competitiveness.
Models and Architectures for Heterogeneous System Design
Tuesday, April 16, 4:10–5:00 pm, 490H Cory Hall, UC Berkeley
Andreas Gerstlauer, University of Texas at Austin
With traditional boundaries between embedded and general-purpose computing blurring, both domains increasingly deal with complex, heterogeneous, software-intensive yet tightly constrained systems. This creates new challenges for modeling, synthesis and component design at the intersection of applications and architectures. For modeling of inherently dynamic behavior, simulations continue to play an important role. We first present an abstract, host-compiled modeling approach that provides a fast and accurate alternative to traditional solutions. By navigating tradeoffs in coarse-grain executions, full-system simulations in close to real time with more than 90% accuracy become possible. Models then provide the basis for automated design space exploration as part of a system compiler that can automatically synthesize parallel application models onto optimized hardware/software platforms. Finally, we will discuss results on co-design of novel domain-specific components for future heterogeneous platforms. We specifically explore fundamental tradeoffs between flexibility and specialization, and between error acceptance and energy in the design of high-performance yet low-power processors for linear algebra and signal processing applications.
A Scalable Adaptive Spatial Discretization for Incompressible Flow using Overset Cartesian Grids
Wednesday, April 17, 11:00 am–12:00 pm, 50F-1647
Elliott English, Stanford University
Adaptive spatial discretizations are important for many computational fluid dynamics problems with large domains and many regions of interest. Methods that use unstructured discretizations are popular for their conceptual simplicity and ability to conform to fluid-structure interfaces. However, these methods often have issues with cache coherency, inaccurate numerical derivatives and domain decomposition. Alternatively, structured methods avoid these issues by parametrically placing degrees of freedom and using simple numerical stencils. However, these methods generally lack adaptivity. Block Cartesian methods such as Adaptive Mesh Refinement (AMR) add adaptivity to structured approaches using axis-aligned Cartesian grid patches. However, due to the large number of patches required to rasterize non-axis-aligned features, AMR approaches often suffer from the same problems as purely unstructured approaches. I present an overset (Chimera) grid method which utilizes arbitrarily rotated and translated Cartesian grid patches. A second order accurate semi-Lagrangian approach is used for advection. The pressure Poisson equation is discretized by patching the grids together using a Voronoi diagram giving a second order accurate SPD linear system. This approach allows us to efficiently solve two-way coupled fluid-structure interaction problems with free surfaces and features at extremely disparate scales not possible with previous approaches.
Dissertation Talk: Avoiding Communication in Dense Linear Algebra
Wednesday, April 17, 12:10–1:00 pm, 380 Soda Hall, UC Berkeley
Grey Ballard, UC Berkeley
The running time of an algorithm depends not only on the amount of computation it performs, but also on its communication requirements (i.e., how much data it moves up and down the memory hierarchy and between processors). In many cases, we can significantly decrease the running times of numerical computations by reformulating existing algorithms to avoid communication.
I will discuss communication lower bounds we have proved for a wide class of algorithms in numerical linear algebra and whether known algorithms achieve these lower bounds. In several cases where gaps exist between algorithms and lower bounds, we have developed new algorithms that communicate asymptotically less than previous ones and outperform them on a variety of platforms. These include algorithms for solving linear systems, least squares problems, eigenvalue problems, and parallelization of Strassen's matrix multiplication algorithm.
Link of the Week: Mathematicians Predict the Future with Data from the Past
Peter Turchin, a professor at the University of Connecticut, is the driving force behind a field called "cliodynamics," where scientists and mathematicians analyze history in the hopes of finding patterns they can then use to predict the future. It’s named after Clio, the Greek muse of history.
These academics have the same goals as other historians — "We start with questions that historians have asked for all of history," Turchin says. "For example: Why do civilizations collapse?" — but they seek to answer these questions quite differently. They use math rather than mere language. According to Turchin, unless something changes, we’re due for a wave of widespread violence in about 2020, including riots and terrorism. Read more.