From EGNg at lbl.gov Wed Jan 9 09:28:05 2008 From: EGNg at lbl.gov (Esmond G. Ng) Date: Wed Jan 9 09:31:55 2008 Subject: [BANANA] Berkeley Lab - Scientific Computing Seminar - Friday, January 18, 2008 Message-ID: <47850425.8010104@lbl.gov> Berkeley Lab - Scientific Computing Seminar Date: Friday, January 18, 2008 Time: 1:00pm-2:00pm Location: Building 50A, 5132 Conference Room Seminar Speaker: Chandrika Kamath Center for Applied Scientific Computing Lawrence Livermore National Laboratory Title: Scientific Data Mining: Challenges at the Petascale Abstract: The data from scientific simulations, observations, and experiments is now being measured in terabytes and will soon reach the petabyte regime. The size of the data, as well as its complexity, make it difficult to find useful information in the data. This is of course disconcerting to scientists who wonder about the science still undiscovered in the data. The Sapphire scientific data mining project at Lawrence Livermore National Laboratory (https://computation.llnl.gov/casc/sapphire) has been addressing this concern by applying data mining techniques to problems ranging in size from a few megabytes to a hundred terabytes in a variety of domains. Using example problems from astronomy, fluid mixing, remote sensing, and experimental physics, I will describe our solution approaches and discuss some of the challenges we have encountered in mining these datasets. Sponsor of Seminar: Arie Shoshani From saunders at stanford.edu Mon Jan 14 10:38:33 2008 From: saunders at stanford.edu (Michael A. Saunders) Date: Mon Jan 14 10:41:48 2008 Subject: [BANANA] LA/Opt Seminar on Wednesday (Victor Pereyra) Message-ID: Linear Algebra and Optimization Seminar (CME510) iCME, Stanford University 4:15pm Wed Jan 16, 2008 Rm 317 Wallenberg Hall (Bldg 160) (Same building as last quarter, but different room) FAST ACOUSTIC WAVE PROPAGATION SIMULATION BY MODEL ORDER REDUCTION Victor Pereyra Weidlinger Associates Mountain View, CA 3DGeo Inc, Santa Clara, CA and CSRC, San Diego State Univ. We will describe the ideas behind a method for reducing drastically the number of degrees of freedom in wave propagation simulations by the method of Proper Orthogonal Decomposition. The basic tenets are: 1. It is necessary to solve many related wave propagation problems, such as in Oil Exploration Earth Imaging. 2. It is feasible to use full fidelity solvers to generate snapshots of the field at appropriate times. 3. It is feasible to calculate the Singular Value Decomposition of the resulting matrix of snapshots. In a pre-processing stage one (or several, but few) full fidelity calculations (expensive) are performed. A matrix of snapshots is created that may have many millions of rows (field variables), but only a few hundred columns (snapshots) and an SVD is calculated. By introducing an appropriate threshold, we choose a basis containing the left singular vectors associated with the significant singular values. With this basis and a Ritz-Galerkin collocation approach a set of time-dependent coefficients is derived by solving a small set of Ordinary Differential Equations. This completes the calculation for the reduced system. We show some numerical results that give an idea of the quality of the approximate solutions obtained by this method and the levels of data compression that arise. Victor Pereyra Weidlinger Associates Inc. 399 El Camino Real, Suite 200 Mountain View, CA 94040 Voice (650) 230-0210 FAX (650) 230-0209 www.wai.com http://homepage.mac.com/vpereyra/pereyra-bio.html/ From saunders at stanford.edu Mon Jan 21 13:46:28 2008 From: saunders at stanford.edu (Michael A. Saunders) Date: Mon Jan 21 13:51:09 2008 Subject: [BANANA] LA/Opt Seminar on Wednesday (Michael Saunders) Message-ID: Linear Algebra and Optimization Seminar (CME510) iCME, Stanford University http://icme.stanford.edu/seminars/seminars.php 4:15pm Wed Jan 23, 2008 Rm 317 Wallenberg Hall (Bldg 160) (Same building as last quarter, but different room) COMPUTING SPARSE PAGERANK VECTORS BY BASIS PURSUIT Michael Saunders SOL, Stanford University http://www.stanford.edu/~saunders/ Many applications involve linear systems Ax ~= b in which an approximate solution x is required to be sparse. Lasso and Basis Pursuit DeNoising were developed for this purpose. They balance the 1-norm of x against the size of the residual, and they allow A to be rectangular. Related procedures include LARS, Homotopy, and the Dantzig Selector. The PageRank eigenvector problem involves a square system Ax = b in which x is sometimes naturally sparse (depending on b). We conduct an empirical study of BPDN in this context. We use an active-set method designed for the dual of the BPDN problem, and find that it tends to extract the important elements of x in a greedy fashion. The test examples come from the Stanford and Berkeley webs. Acknowledgements to Sou-Cheng Choi, Michael Friedlander, David Gleich, and Holly Jin. From EGNg at lbl.gov Thu Jan 24 21:17:57 2008 From: EGNg at lbl.gov (Esmond G. Ng) Date: Thu Jan 24 21:19:08 2008 Subject: [BANANA] Berkeley Lab - Scientific Computing Seminar - Monday, January 28, 2008 Message-ID: <47997105.2030900@lbl.gov> Berkeley Lab - Scientific Computing Seminar Date: Monday, January 28, 2008 Time: 9:45am-10:45am Location: Building 50F, 1647 Conference Room Seminar Speaker: Ichitaro Yamazaki Department of Computer Science University of California, Davis Title: High-Quality Preconditioning Techniques for the Multi-Length-Scale Hybrid Quantum Monte Carlo Simulation of the Hubbard Model Abstract: The hybrid quantum Monte Carlo (HQMC) simulation of the Hubbard model is a powerful tool for studying the electron interactions that characterize the essential properties of materials, such as magnetism and superconductivity. The bottleneck of the HQMC simulation is on the repeated solutions of the underlying multi-length-scale symmetric positive definite (SPD) linear systems of equations. The traditional direct method to solve the linear system is effective for a small number of electrons, but its computational cost scales cubically with the number of electrons. Subsequently, the HQMC simulation with the direct method is limited to hundreds of electrons. In this talk, we present a new preconditioned iterative solver that demonstrates an optimal linear-scaling complexity of the HQMC simulation for moderately-correlated materials. This allows us to conduct simulations with thousands of electrons. The success of the iterative solver relies on a new high-quality incomplete Cholesky (IC)-based preconditioning technique that we propose. The talk will focus on the analysis and implementation of the preconditioning technique for the solution of multi-length-scale SPD systems. We will also introduce a new software package, ICPACK, to provide a uniform interface for IC-based preconditioners. Sponsor of Seminar: Sherry Li From saunders at stanford.edu Tue Jan 29 09:57:37 2008 From: saunders at stanford.edu (Michael A. Saunders) Date: Tue Jan 29 10:03:53 2008 Subject: [BANANA] Stanford OR seminar - Wed Jan 30 - Emilie Danna Message-ID: Dear Stanford Linear Algebra and Optimization folk, This week we have an Operations Research seminar at 4:30. http://or.stanford.edu/ Michael ---------- Forwarded message ---------- Date: Thu, 24 Jan 2008 15:49:57 -0800 From: Hamid Nazerzadeh To: or-seminars@lists.stanford.edu Subject: OR seminars - Wednesday (Jan 30) - Emilie Danna Generating Multiple Solutions for Mixed Integer Programming Problems Emilie Danna ILOG Inc (Sunnyvale HQ) edanna@ilog.com http://www.ilog.com Wednesday, January 30, 2008 4:30 - 5:30 PM Terman Engineering Center, Room 453 As mixed integer programming (MIP) problems become easier to solve in practice, they are used in a growing number of applications where producing a unique optimal solution is often not enough to answer the underlying business problem. Examples include problems where some optimization criteria or some constraints are difficult to model, or where multiple solutions are wanted for quick solution repair in case of model changes. In this talk, we address the problem of effectively generating multiple solutions for the same model, concentrating on optimal and near-optimal solutions. We present a modification of the branch-and-cut algorithm that allows it to produce multiple solutions. We then show that it significantly outperforms previously known algorithms in terms of the speed to generate multiple solutions, while providing an acceptable level of diversity in the solutions produced. From EGNg at lbl.gov Tue Jan 29 16:31:59 2008 From: EGNg at lbl.gov (Esmond G. Ng) Date: Tue Jan 29 16:33:12 2008 Subject: [BANANA] Berkeley Lab - Scientific Computing Seminar - Monday, February 4, 2008 Message-ID: <479FC57F.1030001@lbl.gov> Berkeley Lab - Scientific Computing Seminar Date: Monday, February 4, 2008 Time: 11:00am-12:00pm Location: Building 50B, 4205 Conference Room Seminar Speaker: Karl Fuerlinger Innovative Computing Laboratory Department of Computer Science University of Tennessee, Knoxville Title: Profiling and Incremental Profiling of OpenMP Applications Abstract: Profiling is often the method of choice for performance analysis of parallel applications due to its low overhead and easily comprehensible results. However, a disadvantage of profiling is the loss of temporal information that makes it impossible to causally relate performance phenomena to events that happened prior or later during execution. The talk presents a simple but useful profiling tool for OpenMP applications, ompP, and describes its utility for overhead and scalability analysis. We also present techniques to add temporal dimension to profiling data by incrementally capturing profiles during the runtime of the application and discuss the insights that can be gained from this type of performance data. Application examples come from the SPEC OpenMP benchmark suite. Sponsor of Seminar: David Skinner From EGNg at lbl.gov Tue Jan 29 16:34:08 2008 From: EGNg at lbl.gov (Esmond G. Ng) Date: Tue Jan 29 16:35:33 2008 Subject: [BANANA] Berkeley Lab - Scientific Computing Seminar - Tuesday, February 5, 2008 Message-ID: <479FC600.2060206@lbl.gov> Berkeley Lab - Scientific Computing Seminar Date: Tuesday, February 5, 2008 Time: 11:00am-12:00pm Location: Building 50F, 1647 Conference Room Seminar Speaker: Erika Fuentes Department of Computer Science University of Tennessee, Knoxville Title: Statistical Learning and Data Mining Techniques for Algorithm Selection for Solving Sparse Linear Systems Abstract: There are many applications and problems in science and engineering that require large-scale numerical simulations and computations. The issue of choosing an appropriate method to solve these problems is very common, however it is not a trivial one, principally because this decision is most of the times too hard for humans to make, or certain degree of expertise and knowledge in the particular discipline, or in mathematics, are required. Thus, the development of a methodology that can facilitate or automate this process and helps to understand the problem, would be of great interest and help. The proposal is to utilize various statistically based machine-learning and data mining techniques to analyze and automate the process of choosing an appropriate numerical algorithm for solving a specific set of problems based on their individual properties. Sponsor of Seminar: David Skinner From EGNg at lbl.gov Tue Jan 29 16:36:18 2008 From: EGNg at lbl.gov (Esmond G. Ng) Date: Tue Jan 29 16:37:34 2008 Subject: [BANANA] Berkeley Lab - Scientific Computing Seminar - Wednesday, February 6, 2008 Message-ID: <479FC682.6010506@lbl.gov> Berkeley Lab - Scientific Computing Seminar Date: Wednesday, February 6, 2008 Time: 11:00am-12:00pm Location: Building 50F, 1647 Conference Room Seminar Speaker: Rong Ge Scalable Performance Lab Department of Computer Science Virginia Polytechnic Institute and State University Title: Theories and Techniques for Efficient High-End Computing Abstract: As large-scale computing systems grow tremendously in size and capacity, improving power and performance efficiency becomes a compelling issue. Today it is common for a supercomputer to consume several megawatts of electric power but deliver only 10-15% of its peak performance for average applications. Such power consumption not only costs millions of dollars annually but also dissipates enormous heat that reduces system reliability and productivity. To address these issues, I have developed theories to model the performance and power in high-end computing systems as well as techniques to optimize power and performance efficiency. In this talk, I will present these theories and techniques, yet focus on the quantitative communication performance models (lognP and log3P) and their usage in improving high-end computing efficiency. Compared to previous models, these models explicitly quantify the cost of memory accesses and middleware communications in distributed systems, and thus provide more accurate performance prediction. Moreover, these models aid algorithm designs that improve performance and efficiency. Results show algorithms designed using the lognP and log3P models can outperform those designed by previous models and reduces execution time by up to 59%. Sponsor of Seminar: David Skinner From EGNg at lbl.gov Tue Jan 29 16:38:14 2008 From: EGNg at lbl.gov (Esmond G. Ng) Date: Tue Jan 29 16:39:31 2008 Subject: [BANANA] Berkeley Lab - Scientific Computing Seminar - Thursday, February 7, 2008 Message-ID: <479FC6F6.601@lbl.gov> Berkeley Lab - Scientific Computing Seminar Date: Thursday, February 7, 2008 Time: 11:00am-12:00pm Location: Building 50F, 1647 Conference Room Seminar Speaker: Yunrong Zhu Department of Mathematics Penn State University Title: Robust Multilevel Preconditioners for Problems with Strongly Discontinuous Coefficients Abstract: Although there are vast literature of multilevel and domain decomposition (DD) methods for the finite element discretization of elliptic (H(grad)) systems, it remains an open question that how to make these efficient solvers convergence (nearly) uniformly for the H(grad) systems with strongly discontinuous coefficients. Recently, we proved that the multilevel and DD preconditioners lead to a nearly uniform convergent preconditioned conjugate gradient methods. In this talk, I will present the theoretical and numerical justification of these results. As applications, I will also present the auxiliary space preconditioners (Hiptmair and Xu 2007) for H(curl) and H(div) systems in the compatible discretization framework, which reduce the Maxwell equations and Mixed formulation of elliptic equations into solving several H(grad) equations. I will give some numerical results for H(div) system and its application. Sponsor of Seminar: Sherry Li