Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: Cosmin Petra Project Name: stoch_prog Division: MCS Project title: Stochastic Programming and Application for Energy Systems Associated funding: DOE-ASCR: Mathematics of Petascale Data Program. Scalable Algorithms for Gaussian Processes. Mihai Anitescu, PI. Other Systems: We have won an INCITE Award of 10 million hours on BG/P. Science: This project is aimed at developing scalable methods and implementations for the solution of stochastic optimization problems on massively parallel computing platforms. On the application side we focus on the optimization of power grid in the presence of uncertainty. Our objective is to illustrate how the use of forecast information, accurate physics and high performance computing reduce the operating costs. Project description: In the previous years, we obtained an unprecedented level of parallelization in solving stochastic optimization problems by exploiting the half-arrow shaped structure these problems have and using Elemental to reduce the computational bottleneck associated with the first stage dense Schur complement system. We also have a hybrid, MPI+SMP implementations. The intranode parallelization is obtained via OpenMP. Our C++ code, PIPS, solved a stochastic economic dispatch problem with close to 2 billion variables using 131072 cores on Intrepid with a strong scaling efficiency of more than 90%. A paper on these findings has been accepted to SC11. The final version can be found here: http://www.mcs.anl.gov/~anitescu/PUBLICATIONS/2011/miles-2011-pips-SC.pdf Currently, we are adding capabilities for integer variables. An example of such variables are the binary (integer) variables representing switches and the “on/off” status of generators (a.k.a unit commitment) present in the power grid optimization models. A first step in the direction of adding capabilities for integer variables was the implementation of PIPS-S, a parallel SIMPLEX implementation for stochastic linear programming problems. To our knowledge, PIPS-S is the first simplex implementation for HPC platforms. The runs on FUSION showed a good strong scaling, up to 50%. A full report on the implementation details and the performance on Fusion and Intrepid was submitted to the Journal of Computational Optimization and Applications. The ANL report can be found at http://www.mcs.anl.gov/~petra/papers/pipss.pdf . Integer optimization problems have a combinatorial nature, and are often solved by recursively exploring the finite but exponentially large binary solution space using an approach called branch and bound. However, parallel branch-and-bound has not previously been used to solve problems so large that the continuous relaxation must be decomposed and solved in parallel. Now we can efficiently solve the relaxations using PIPS-S and we will pursue the implementation of a parallel branch-and-bound code. The requested hours on Fusion will be used in the development phases of the code to run small and medium scale computations. Large scale runs and studies will be done on BG/P or Q platforms Project URL: http://www.mcs.anl.gov/~petra/pips.html Current FY Hours Used: undetermined amount New FY Requested allocation: 140000 Q1: 10000 Q2: 10000 Q3: 40000 Q4: 80000 Justification: Thank You, The LCRC Accounts System