LANS Informal Seminar: LANS Summer Students

Seminar Title: SASSy – Part 1
Speaker: LANS Summer Students,

Date/Time: 2015-07-31 14:30
Location: Building 240 rm 1404-1405

Description:
Speaker list:
2:30 PM Amit Roy Paul Hovland
2:45 PM Devashri Nagarkar Paul Hovland
3:00 PM Harikrishan Sreekumaran Todd Munson
3:15 PM John Mangeri Dimitry Karpeyev, Barry Smith

3:30 PM BREAK

3:45 PM John O’Sullivan Barry Smith
4:00 PM Mahesh Narayanamurthi Sri Krishna Narayanan
4:15 PM Christian Tjandraatmadja Victor Zavala
4:30 PM Wanting Xu Mihai Anitescu

Abstracts:

Session 1 2:30-3:30 PM
Session Chair: Kamil Khan

2:30-2:45 PM
Student: Amit Roy
Supervisor: Paul Hovland
Title: Exploiting performance portability in search algorithms for auto tuning
Abstract: Autotuning for code performance deals with finding the best implementation of an application by orchetrating various hardware and software knobs that affect the performance on a given machine. Many autotuners adopt various search techniques to efficiently find the best configuration. A major criticism of autotuning is that each application needs to be retuned when moving from one machine to another machine, which is a computationally expensive process. We present a machine-learning-based approach to build surrogate performance model from the autotuning results obtained from one machine to speed up the search on another machine. The proposed approach speeds up the autotuning search by 2x to 5x on a variety of modern architectures.

2:45-3:00 PM
Student: Devashri Nagarkar
Supervisor: Paul Hovland
Title: Python Testsuite for Automatic Differentiation Abstract: Automatic Differentiation (AD) is a tool that differentiates functions in computer programs. It relies on the concept that even the most complex function is comprised of simple functions such as addition, subtraction, multiplication or division, therefore can be derived using the chain rules. AD saves researchers time and is compatible with programs that symbolic differentiation and numerical differentiation cannot handle. AD has been implemented in a variety of languages; the most thorough implementation exists in C and FORTRAN. There is a large gap in AD when it comes to Python: an increasingly popular language. The drawback of Python is that is runs slower than languages like C and FORTRAN. To speed up the process of differentiating, functions and drivers can also be implemented using a Python library called Numpy. Numpy is basically C code wrapped in Python. We have compiled 39 mathematical functions in Python and duplicated them using Numpy. We will then use the functions to test the Python implementation of AD after it is ready. After it is created, the AD tool that preserves the advantages of Numpy implementation should be faster, finally creating an effective and fast tool for AD in Python.

3:00-3:15 PM
Student: Harikrishan Sreekumaran
Supervisor: Todd Munson
Title: Bilevel Mixed-Integer Non-linear Programming and Robust Optimization
Abstract: Bilevel programming problems arise naturally in both hierarchical decision making as well as robust optimization models. These models have been used in a variety of engineering applications ranging from network planning and vulnerability analysis to radiation treatment planning for cancer patients. In this talk, we briefly review bilevel nonlinear programming problems(BL NLPs). We present relaxation formulations for mixed integer BL NLPs, which we utilize in an outer-approximation as well as a branch and bound algorithm. We also review bilevel formulations of robust optimization problems and discuss some ideas for solving such problems.

3:15-3:30 PM 
Student: John Mangeri
Supervisor: Dimitry Karpeyev, Barry Smith
Title: FERRET – A scalable code package within the MOOSE/libMesh/PETSc framework for simulation of ferroelectric nanoparticles
Abstract: FERRET, a code package developed within the MOOSE/libMesh/PETSc framework, for simulating ferroelectric materials will be overviewed. These systems, whose primary order parameter is a polarization field that arises from fixed dipoles on the lattice, exhibits direct and nontrivial coupling to elastic deformations. Using the Landau-Ginzburg equations that describe a energy functional that is dependent on specific material symmetry, a time-dependent scheme is developed that accounts for the domain-domain interactions, response to depolarizing and applied field, and elastic boundary conditions on the nanoscale. Recent results of this implementation will be presented and discussed.

BREAK 3:30-3:45 PM

Session 2 3:45-4:45 PM
Session Chair: Kibaek Kim

3:45-4:00 PM
Student: John O’Sullivan
Supervisor: Barry Smith
Title: Upgrading SAWs for PETSc
Abstract: PETSc is a diverse framework of mathematical tools that enables scientific development across multiple spectrums. It uses a program called SAWs, which allows scientists to view their experiments in real-time or change parameters as the experiment progresses. The web aspect was not fully developed. We renovated it to appeal to the user, but realized that scientists working on long experiments could not view past and present data at the same time. To resolve the data set dilemma, I installed a data structure into SAWs which, once enabled in PETSc, would allow users to simply click a button and be able to select and view previous data. Beforehand, SAWs only enabled one path for resources, which hindered development. To address this issue, my sponsor, Barry Smith, implemented multiple resource parameter support in PETSc which allowed me to continue my research in SAWs. The PETSc program I helped develop has created two new features that will aid researchers design computational methods on extremscale computers for complex applications.

4:00-4:15 PM 
Student: Mahesh Narayanamurthi
Supervisor: Sri Hari Krishna Narayanan
Title: A test suite for AD tool advances
Abstract: Derivatives of models are an essential part of any simulation. Finite differences, although not too difficult to implement, can give poor derivative estimates and its cost increases as number of variables times computations per model run. One solution is to hand-code the derivative of the model and use it to compute the derivatives. But this can be tedious to maintain if the model is being constantly revised or is too large with several stages of computation. Automatic differentiation can remove the need for hand-coding the derivatives and simultaneously provide derivates that are usually accurate to within a reasonable multiple of machine epsilon. Tools like OpenAD/F automate the process of generating code for computing the derivatives of large models by performing source transformation of the original model. A suite of mini-apps spanning application of AD to CFD, Electrical Power Transmission, Atmospheric Chemistry and Reservoir Simulation has been put together to test advances of these tools. I’ll also discuss some challenges that I encountered in the process.

4:15-4:30 PM
Student: Christian Tjandraatmadja
Supervisor: Victor Zavala
Title: Experimenting with Decomposition Methods for the Unit Commitment Problem
Abstract: The stochastic unit commitment problem is a classic problem in energy dispatch: which generators should we turn on in order to satisfy a stochastic demand on a power grid? Although the unit commitment problem has been extensively studied for decades, it remains an important challenge to solve large-scale instances in a practical amount of time. The aim of our work is to push towards this direction by experimenting with decomposition methods using DSP, a recently released parallel solver for stochastic mixed-integer problems developed at Argonne by Kibaek Kim and Victor Zavala. In the first part of our work, we extend the algorithms in DSP to support general decomposition methods. We include a Julia interface that allows a user to model a decomposable version of a problem. This enables us (and others) to quickly prototype and test decomposition methods in stochastic mixed-integer programming. In the second part, we seek to tackle requirements of large-scale instances, such as large networks and fine temporal resolution. We use our DSP extension to assess the potential of network decomposition and time decomposition methods.

4:30-4:45 PM
Student: Wanting Xu
Supervisor: Mihai Anitescu
Title: Improving low memory state estimation of weakly constrained 4D-Var with gradient evaluation
Abstract: Data assimilation is the process of estimating the states of a dynamical system with observation data. The underlying states and observables form a hidden Markov model, and the best states can be obtained by minimizing a corresponding likelihood function. We developed a recursive scheme for evaluating the gradient of the object function in a low memory fashion, and use that information in optimization. We demonstrate the improvement of our method over the existing low memory approach by simulation with Burgers’ equation and a linear PDE.