Abstract: A path optimization method is presented here for analysis of travel times of seismic waves.The method is an adaption of the nudged elastic band method to ray tracing where the the path corresponding to minimal travel time is determined. The method is based on a discrete representation of an initial path followed by iterative optimization of the discretization points so as to minimize the integrated time of propagation along the path. The gradient of the travel time with respect to the location of the discretization points is evaluated and used to find the optimal location of the points. An important aspect of the method is an estimation of the tangent to the path at each discretization point and elimination of the component of the gradient along the path during the iterative optimization. The distribution of discretization points along the path is controlled by spring forces acting only in the direction of the path tangent. The method is illustrated on two test problems and performance compared with previously proposed and actively used methods in the field of seismic data inversion.

Abstract: The GOUST method relies on a fast way to identify first order saddle points on multidimensional objective function surfaces [1,2]. Given a local minimum, the method involves farming out several searches for first order saddle points, and then sliding down on the other side to discover new local minima. The system is then advanced to one of the newly discovered minima. In this way, local minima of the objective function are mapped out with a tendency to progress towards the global minimum. A practical application of this approach in global optimization of a few geothermal reservoir model parameters has recently been demonstrated [3]. The difficulty of an optimization problem, however, generally increases exponentially with the number of degrees of freedom and we investigate GOUST’s ability to search for the global minimum using a group of selected test functions. The performance of GOUST is tested as a function of the number of dimensions and compared with various other global optimization methods such as evolutionary algorithms and basin hopping. [1] 'A Dimer Method for Finding Saddle Points on High Dimensional Potential Surfaces Using Only First Derivatives',G. Henkelman and H. Jónsson, J. Chem. Phys., Vol. 111, page 7010 (1999) [2] 'Comparison of methods for finding saddle points without knowledge of the final states’, R. A. Olsen, G. J. Kroes, G. Henkelman, A. Arnaldsson and H. Jónsson, J. Chem. Phys. vol. 121, 9776 (2004). [3] 'Geothermal model calibration using a global minimization algorithm based on finding saddle points as well as minima of the objective function', M. Plasencia, A. Pedersen, A. Arnaldsson, J-C. Berthet and H. Jónsson, Computers and Geosciences 65, 110 (2014)

Abstract: In this work we propose a memory-efficient (cache oblivious) implementation of the Finite Difference Time Domain algorithm. The implementation is based on a recursive space-time decomposition of the mesh update dependency graph into subtasks. The algorithm is suitable for processing large spatial meshes, since, unlike in the traditional layer-by-layer update, its efficiency (number of processed mesh cells per unit time) does not drastically drop with growing total mesh size. Additionally, our implementation allows for concurrent execution of subtasks of different size. Depending on the computer architecture, the scheduling may simultaneously encompass different parallelism levels such as vectorization, multithreading and MPI. Concurrent execution mechanisms are switched on (programmed) for subgraphs reaching some suitable size (rank) in course of recursion. In this presentation we discuss the implementation and analyze the performance of the implemented FDTD algorithm for various computer architectures, including multicore systems and large clusters. We demonstrate the FDTD update performance reaching up to 50% of the estimated CPU peak which is 10-30 times higher than that of the traditional FDTD solvers. We also demonstrate an almost perfect parallel scaling of the implemented solver. We discuss the effect of mesh memory layouts such as Z-curve (Morton order) increasing locality of data or interleaved layouts for vectorized updates.

Abstract: The Consortium for Advanced Simulation of Light Water Reactors (CASL) was established in July 2010 for the modeling and simulation of commercial nuclear reactors. Led by Oak Ridge National Laboratory (ORNL), CASL also includes three major universities, three industry partners, and three other U.S. National Laboratories. In order to deliver advanced simulation capabilities, CASL has developed and deployed VERA, the Virtual Environment for Reactor Applications (VERA), which integrates components for physical phenomena to enable high-fidelity analysis of conditions within nuclear reactors under a wide range of operating conditions. We report on the architecture of the system, why we refer to it as an Environment rather than a Toolkit or Framework, numerical approaches to the coupled nonlinear simulations, and show results produced on large HPC systems such as the 300,000-core NVIDIA GPU-accelerated Cray XK7 Titan system at Oak Ridge National Laboratory.