InTheLoop | 01.13.2014
NERSC Helps BOSS Measure the Universe to One-Percent Accuracy
The Baryon Oscillation Spectroscopic Survey (BOSS) Collaboration announced last week that BOSS has measured the scale of the universe to an accuracy of one percent. This and future measures at this precision are the key to determining the nature of dark energy. Complex computer algorithms were essential for reconciling inherent uncertainties in BOSS's data. "We made thousands of model universes in a computer, and then observed them as BOSS would do and ran our analysis on them to answer the questions of 'What if?' "
By gauging how well their algorithms could analyze model universes, known as "mocks" and based on catalogues of realistic but artificial galaxies, the BOSS team was able to assess and fine-tune the algorithms when they were applied to the real BOSS data. The National Energy Research Scientific Computing Center (NERSC) was critical to the analysis and the creation of the mocks. »Read more.
NERSC Kicks Off 40th Anniversary Celebration With 2014 Calendar
The National Energy Research Scientific Computing Center (NERSC) is kicking off a year-long celebration of its 40th anniversary with a free, downloadable 2014 calendar that takes a unique look back at our supercomputing history. The 2014 calendar features historical images and unique trivia highlighting the people and machines that have made NERSC such a success over the last 40 years.
Kennedy High Graduates Get Hands-On Experience Working in Computing Sciences
From a connection that started between soccer parents, Berkeley Lab's Computing Sciences organization has developed an ongoing outreach program with the IT Academy at Kennedy High School in Richmond. This summer, that connection led to summer jobs for three Kennedy graduates, giving them hands-on experience and a paycheck. »Read more.
CRD's Oliker Co-Edits Special Edition of Journal
Lenny Oliker of CRD recently co-edited a special issue of the International Journal of High Performance Computing Applications. The issue focuses on automated performance tuning (autotuning) techniques. Autotuning is becoming increasingly important as supercomputer designs diverge in response to the challenge of reaching exascale-level performance. »Read more.
SC14 Now Accepting Workshop Submissions, Deadline February 7
Interested in hosting a workshop at the SC14 conference to be held Nov. 16-21 in New Orleans?
SC14 is accepting proposals until Friday, February 7 for independently planned full-, half-, or multi-day workshops. Workshops complement the overall SC technical program by providing focused, in-depth venues for presentations, discussion and interaction in particular subject areas. Workshop proposals will be peer-reviewed with a focus on submissions that will inspire deep and interactive dialogue in topics of interest to the HPC community.
»Get More Information
»Submit a Proposal
»Ask a Question
This Week's Computing Sciences Seminars
Computer Architecture Research at ACS
Monday, January 13, 10am – 11am, Bldg. 50F, Room 1647
Noel Wheeler, Laboratory for Physical Science Advanced Computing Systems University of Maryland
No abstract was submitted for this talk.
Understanding Images of Particle Ensembles in Cryo Electron Microscopy in Terms of Statistical Heterogeneity of the Particles
Tuesday, January 14, 3pm – 4pm, Bldg. 50B, Room 4205
Peter C. Doerschuk, Department of Biomedical Engineering School of Electrical and Computer Engineering Cornell University, Ithaca NY
Cryo electron microscopy provides one image of each of many instances of the biological particle of interest. Algorithms and software for characterizing the 3-D heterogeneity of the particles based on the images through solving maximum likelihood estimation problems are described. Application to Flock House Virus (FHV) and Nudaurelia omega capensis virus (NwV) are described. Current work on complete incorporation of symmetry constraints on the statistics describing the heterogeneity and on understanding the heterogeneity in terms of generative models will be sketched.
Peter Doerschuk received BS, MS, and Ph.D. degrees in electrical engineering from MIT and an MD degree from Harvard Medical School. After a postdoc at MIT he was on the faculty in Electrical and Computer Engineering at Purdue University from 1990 to 2006 and is presently on the faculty in Biomedical Engineering and Electrical and Computer Engineering at Cornell University. His interest is in developing and using ideas in statistical detection and estimation theory to address problems in understanding large biological or medical data sets.
A Model for Expressing Virtual Organization Identity Management
Wednesday, January 15, 1:00pm – 2:30pm, Bldg. 50B, Room 4205
Von Welch, Deputy Director, Indiana University Center for Applied Cybersecurity Research
PI, eXtreme Scale Identity Management Project, Laboratory for Physical Science Advanced Computing Systems, University of Maryland
Over the past two decades, the virtual organization has allowed for increasingly large scientific virtual organizations (VOs) spanning multiple organizations and nations. How identity management, i.e. the management of users and their privileges, in these large VOs is handled has been an ongoing area of applied research. The vision of the eXtreme Scale Identity Management (XSIM) project is to develop a model for expressing different approaches for identity management, with the strengths and ramifications of each, and guidance to VOs which models best fits their needs. This talk will describe XSIM's efforts in the initial stage of this effort, the descriptive model, and its future efforts in developing the guidance to VOs in selecting a identity management model.
Multigrid on the Outside: Restructuring time integration and adaptivity
Thursday, January 16, 2:30pm – 3:30pm, Bldg. 50A, Room 5132
Jed Brown, Mathematics and Computer Science, Argonne National Laboratory
Performance of transient PDE simulation is typically limited by memory bandwidth, insufficient vectorization, and/or communication latency, as well as convergence when implicit solvers are used. We present two new restructurings that alleviate these costs. The first exploits tensor product structure in fully implicit Runge-Kutta methods. The method is unintrusive and naturally implemented using sparse matrices, generalizing multiple right hand sides with the same performance returns. The second restructuring is a multilevel adaptive algorithm for stiff nonsmooth problems in heterogeneous media, such as contact or plasticity. Conventional adaptive methods are not applicable because of the microstructure, but after a small amount of work on the fine grid, our method allows elision of fine-grid visits except in the neighborhood of active nonlinearities. The resulting algorithm is enticing relative to nonlinear domain decomposition methods such as ASPIN due to reduced overall work and because load balancing is tractable since it is a one-stage method.