Seminar Details:

LANS Informal Seminar
"Scalable Stochastic Optimization of Complex Energy Systems"

DATE: July 6, 2011

TIME: 15:00:00 - 16:00:00
SPEAKER: Miles Lubin, Predoctoral Appointee, MCS
LOCATION: Bldg 240 Conference Center 1404-1405, Argonne National Laboratory

We present a scalable approach and implementation for solving stochastic programming problems, with application to the optimization of complex energy systems under uncertainty. Stochastic programming incorporates a model of uncertainty about future events, which in our case relates to the output of highly variable renewable energy sources such as wind. The Sample Average Approximation (SAA) is used to obtain a very large deterministic problem, which necessitates the use of parallel computing to solve.

Our code, PIPS, is based on primal-dual interior-point methods (IPMs) and uses a classical Schur-complement technique to obtain a sample- or scenario-based decomposition at the linear algebra level of IPMs. We review the Schur-complement decomposition and present our novel method for solving the Schur-complement system using distributed dense linear algebra, allowing, for the first time, problems with up to 100,000 first-stage variables (and ~2 billion total variables) to be solved efficiently. We discuss our experience with implementing a hybrid parallel model and porting PIPS to Blue Gene/P, where we obtained strong scaling efficiency of 90%+ on 32 racks (131,072 cores) of Intrepid. Our application demonstrates the potential of efficiently using High Performance Computing resources for mathematical optimization.


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