applied math seminar
Eric T. Phipps (presenter) and Marta D’Elia Center for Computing Research, Sandia National Laboratories, Albuquerque, New Mexico
Monday, April 23, 2018 -
3:30pm to 4:30pm
Title: Improving Sampling-based Uncertainty Quantification Performance Through Embedded Ensemble Propagation
A key component of computational uncertainty quantification is the forward propagation of uncertainties in simulation input data to output quantities of interest. Typical approaches involve repeated sampling of the simulation over the uncertain input data, and can require numerous samples when accurately propagating uncertainties from large numbers of sources. Often simulation processes from sample to sample are similar and much of the data generated from each sample evaluation could be reused. In this talk, we explore a new method for implementing sampling methods that simultaneously propagates groups of samples together in an embedded fashion, which we call embedded ensemble propagation . We show how this approach exploits properties of modern computer architectures to improve performance by enabling reuse between samples, reducing memory bandwidth requirements, improving memory access patterns, improving opportunities for fine-grained parallelization, and reducing communication costs. We describe a software technique for implementing embedded ensemble propagation based on the use of C++ templates, and demonstrate improved performance for the approach when applied to model diffusion problems on a variety of contemporary architectures. A challenge with this method however is ensemble-divergence, whereby different samples within an ensemble choose different code paths. This can reduce the effectiveness of the method and increase computational cost. Therefore grouping samples together to minimize this divergence is paramount in making the method effective for challenging computational simulations. We also present several grouping approaches [1, 2] that attempt to minimize this divergence through surrogate models of ensemble computational cost. These approaches are developed within the context of locally adaptive stochastic collocation methods and are applied to highly anisotropic diffusion problems where computational cost is driven by the number of (preconditioned) linear solver iterations, which vary widely from sample to sample.  M. D’Elia, H. C. Edwards, J. Hu, E. Phipps, and S. Rajamanickam. Ensemble Grouping Strategies for Embedded Stochastic Collocation Methods Applied to Anisotropic Diffusion Problems. SIAM/ASA Journal on Uncertainty Quantification, 6(1):87–117, Jan. 2018.  M. D’Elia, E. Phipps, A. Rushdi, and M. Ebeida. Surrogate-based ensemble grouping strategies for embedded sampling-based uncertainty quantification. Submitted to SIAM/ASA Journal on Uncertainty Quantification, 2017. https://arxiv.org/abs/1705.02003.  E. Phipps, M. D’Elia, H. C. Edwards, M. Hoemmen, J. Hu, and S. Rajamanickam. Embedded ensemble propagation for improving performance, portability, and scalability of uncertainty quantification on emerging computational architectures. SIAM Journal on Scientific Computing, 39(2):C162–C193, 2017.
Eric joined Sandia National Laboratories in September 2002 and is currently a principal member of the technical staff in the Optimization and Uncertainty Quantification Department. Eric's research focuses on developing new capabilities for predictive simulation and analysis in Sandia’s large-scale parallel application codes using techniques based on automatic differentiation and template-based generic programming. His work has recently emphasized developing tools and techniques relevant to emerging extreme-scale computer architectures. He is the lead developer for several related software packages in Trilinos including the Sacado automatic differentiation, Stokhos embedded uncertainty quantification, and LOCA continuation/bifurcation analysis packages. See https://cfwebprod.sandia.gov/cfdocs/CompResearch/templates/insert/profile.cfm?etphipp for selected publications.
Contact Name: Pavel Lushnikov