MACSER Investigator Delivers Keynote Address on Scalable Stochastic Programming

Mihai Anitescu gave a plenary presentation March 22, 2018, at the 7th international conference on High Performance Scientific Computing (HPSC) in Hanoi, Vietnam. His presentation, titled “Scalable Stochastic Programming for Energy Systems,” focused on the challenges of modeling and optimization of complex processes in the presence of uncertainty.

Anitescu, a senior computational mathematician in Argonne’s Mathematics and Computer Science Division and lead of the MACSER project, began by describing the computational difficulties stochastic programmers face in trying to make decisions in the present while incorporating a model of uncertainty about future events.

“As the number of scenarios becomes large and the complexity of the system and planning horizons increase, parallel computing becomes essential,” Anitescu said.

Anitescu discussed a series of technical ideas for achieving scalability. These include a parallel MPI+OpenMP interior-point solver for stochastic linear programs and convex quadratic programs, a parallel MPI implementation of the revised simplex method, and a parallel MPI interior-point method for structured nonlinear programs. Each of these solvers has been implemented in the PIPS suite for stochastic programming, developed at Argonne National Laboratory and available for download on GitHub at https://github.com/Argonne-National-Laboratory/PIPS.

To demonstrate the effectiveness of the solvers, Anitescu and his colleagues applied the PIPS suite to large-scale stochastic energy dispatch problems. The experiments were run on massively parallel supercomputers with more than 100,000-way parallelism.

“We demonstrated for the first time at this scale that such problems can benefit from existing and emerging very high-performance computing architectures,” Anitescu said. He acknowledged that numerous challenges remain – longer time horizons, more frequent temporal decisions, larger spatial networks – but he expressed confidence that with algorithmic improvements such challenges can be met, saving the energy industry hundreds of gigawatts.