Seminar Details:

LANS Informal Seminar
"SASSy (Summer Argonne Students Symposium) Part 2"

DATE: August 14, 2013

TIME: 13:00:00 - 17:00:00
SPEAKER: LANS Summer Students
LOCATION: Building 240 Room 1406-1407, Argonne National Laboratory


+ Lihua (Tom) Li (Supervisor: Jie Chen/Mihai Anitescu)
1:00 - 1:15 PM
Title:Parallel Block Conjugate Gradient Algorithm for Multilevel Toeplitz Systems
Abstract: Multilevel Toeplitz linear systems appear in a wide range of scientific and engineering applications. While several fast direct solvers exist for the basic 1-level Toeplitz matrices, in the multilevel case an iterative solver provides the most general and practical solution. We implemented the block conjugate gradient algorithm using MPI calls. Techniques that enable efficient computation are also introduced: (1) Adoption of computation/communication overlapping scheme using non-blocking collective calls; (2) flexible configuration for multi-dimension processor grid partitioning. (3) elimination of Allreduce MPI calls to avoid unnecessary synchronization.

+ Elias Nino-Ruiz (Supervisor: Cosmin Petra)
1:15 - 1:30 PM
Title: Low-sample Covariance Estimation For Power Grid Models
Abstract: We report on the use of estimators in order to approach the correlation in time between wind farms in the State of Illinois when the number of components $p$ of the random vector is much larger than the sample size $n$. Initially, we review some estimators from the literature and we note, under the condition $p \gg n$, the best approximation is given by the shrinkage based estimators. Secondly, these kind of estimators are utilized in order to estimate the correlation in time between wind farms in the state of Illinois. Lastly, the results are validated empirically. We note that, the results are consistent with the wind flow directions, the proximity of the points and the wind speed at the given wind farm locations.

+ Yankai Cao (Supervisor: Victor Zavala)
1:30 - 1:45 PM
Title: Title Clustering-Based Interior-Point Strategies for Stochastic Programs
Abstract: We present interior-point strategies for convex stochastic programs in which compressed Karush-Kuhn-Tucher(KKT) systems obtained by clustering block scenarios are used as pre-conditioners. We demonstrate that an order of magnitude speedup is possible compared with solving full space KKT systems for our network problems.

+ Nick Horelik (Supervisor: Andrew Siegel)
1:45 - 2:00 PM
Title: Implementing Domain Decomposition in the OpenMC Monte Carlo Neutron Transport Code
Abstract: Monte Carlo particle transport methods are considered the "gold-standard" for nuclear simulations with respect to accuracy, being pursued as one path to efficient high-fidelity exascale computing at CESAR (Center for Exascale Simulation of Advanced Reactors). The method presents enormous parallelizability, but memory constraints restrict the level of detail that can be handled in the problem sizes of interest. This talk will cover an overview of these challenges and present a robust domain decomposition algorithm that promises to alleviate them, as implemented in the OpenMC Monte Carlo code this summer at ANL.

+ Anton Bovin (Supervisor: Jean Utke)
2:00 - 2:15 PM
Title: Reduction Operations in AdjoinableMPI
Abstract: Differentiation of mathematical expressions has many uses in scientific computing; of particular interest to us is the case of scalar-valued functions, for which the Jacobian reduces to a gradient. We employ the AD (algorithmic differentiation) reverse/adjoint mode to differentiate such functions by associating variables with derivative values and performing a trace of the computation. Parallelizing the evaluation and subsequent differentiation requires augmentation of MPI routines (implemented in the AdjoinableMPI wrapper library) with functionality to perform the appropriate operations during the reverse sweep. Currently AdjoinableMPI implements gather/scatter and blocking/nonblocking send/receive routines; my project has focused on extending AdjoinableMPI to include broadcast/reduce functions as well as support for user-defined types and reduction operations. Primary challenges include managing the indirection between active variables and their real-number values, as well as ensuring the tool-independence of the wrapper library.

-- 15 Minute Break --

+ Jack Burrell (Supervisor: Mihai Anitescu)
2:30 - 2:45 PM
Title: Electrochemical Materials Discovery: From Performance to Stability
Abstract: I will be talking about how a battery works, the problems, and why we want to model the batteries on a computer instead of making them in a lab.

+ Devin O'Connor (Supervisor: Mihai Anitescu)
2:45 - 3:00 PM
Title:Phase-Field Modeling of Lithium Intercalation and Fracture in LiFePO4 Electrodes
Abstract: When a lithium-ion battery is charged or discharged lithium ions are either removed or inserted (intercalated), respectively, into the electrode. In silicon-based anodes this process can cause a huge volume expansion in the electrode. During lithium removal or insertion the electrode can also undergo a phase transformation, as seen in LiFePO4 cathodes, which can cause large misfit strains at phase boundaries. In both cases, fractures in electrodes and subsequent capacity fading have been seen. To be able to design better lithium-ion batteries it is important to be able to understand how these cracks arise and how to mitigate them. In this talk I will present a coupled multi-physics model that we have been developing this summer that simulates the phase transformation and fracture that has been seen in LiFePO4 electrodes. Preliminary results have shown that the crack has an influence on the lithium ion mobility around the region of the crack tip and this build-up of lithium could lead to crack propagation and electrode failure.

+ Jesse Lopez (Supervisor: Jed Brown)
3:00 - 3:15 PM
Title: Scaling 3D SELFE: A semi-implicit finite element ocean model
Abstract: Here we present the experiences and results of an attempt to improve the scalability of an finite element based 3D circulation model used by oceanographers and environmental scientists for estuarine and coastal studies. Through the integration of profiling into the development cycle bottlenecks impeding scaling were identified and ameliorated or eliminated. Solutions included selective refactoring, avoiding disk intensive IO, and the porting of a linear solver from custom code to the PETSc library.

+ Patrick Sanan (Supervisor: Barry Smith/Jed Brown)
3:15 - 3:30 PM
Title: Structured Integrators in PETSc
Abstract: To solve initial value problems for ordinary differential equations (ODEs), researchers have developed a broad range of integrators (time-steppers) which exploit "structure". This structure often takes the form of a splitting of the righthand side of the ODE dx/dt = f(x) + g(x). I'll present a preliminary interface for partitioned right hand sides in PETSc's TS package and show some toy examples for splittings used by structured integrators. Of particular interest for large scale physical simulations are systems with two well-separated timescales: "slow" behavior which captures macroscopic observables and "fast" behavior which influences the slow behavior, yet need not be fully resolved. The fast behavior can make a single-scale approach intractable due to stiffness, oscillatory behavior, and/or high dimensionality at the fine scale. After presenting preliminary implementations of multiscale integrators and a potential application in complex fluid simulation, I'll end with a request for comment on the interface and applications, as I still have a month to improve things!

+ Mu Wang (Supervisor: Jean Utke)
3:30 - 3:45 PM
Title: The Design and Implementation of Reverse Mode in Rapsodia
Abstract: Automatic Differentiation (AD) is a technology for analytically evaluating the derivatives of a function specified by a computer program. Rapsodia is an AD tool which combines operator overloading and code generation. The generated code makes further optimizations taken by compilers possible. This summer we added the reverse mode in Rapsodia with some design innovations such as pre-accumulation and reuse the deallocated stack memory.

+ Cao (Eric) Ni (Supervisor: Stefan Wild)
3:45 - 4:00 PM
Title: Zero-Order Optimization with Unrelaxable Integer Constraints
Abstract: We review various algorithms for solving nonlinear optimization problems with integer constraints when derivatives of the objective and relaxations of the integer constraints are both unavailable. We will discuss one branch-and-bound strategy that guarantees global optimality when the objective function is convex.


Please send questions or suggestions to Jeffrey Larson: jmlarson at anl dot gov.