SC²S Colloquium - March 30, 2016
|Date:||March 30, 2016|
|Time:||3:00 pm, s.t.|
Aleksandra Pachalieva: Parallel, Dynamically Adaptive 2.5D Porous Media Flow Simulation on Xeon Phi Architectures
sam(oa)2 is a fast and memory efficient framework for the solution of PDEs based on Sierpinski curve traversals. The hybrid MPI+OpenMP parallelization of the code utilizes data parallelism in distributed and shared memory, which is a premise to achieve good performance on HPC systems. sam(oa)2 contains a 2.5D porous media flow scenario suitable for vectorization with its regular grid in the vertical layers that makes it a good candidate for running on the Intel XeonPhi architecture. Further optimization techniques are proposed like vectorization, taking advantage of the 512-bit SIMD vector support, and load balancing, essential for the scalability due to the problem specific remeshing and varying computational load in symmetric mode. The performance of the framework is tested on the SuperMIC supercomputer using time-based load balancing.
Florian Klemt: Sparse grids with data-aware hierarchical subspaces for regression
The goal of this bachelors thesis is to construct and implement a median based sparse grid and apply it for regression problems. By basing the grid on medians we obtain a data-aware structure. For non-uniform datasets this leads to better regression results than a regression based on regular sparse grids. We test and validate the method and compare it against regular sparse grids. In order to do that we wrote a program in the programming language Python consisting of two parts, one constructing the median based sparse grid on a given dataset and one using it to compute regression on the same dataset. For testing purposes we used a lot of artificially generated datasets, to show the basic improvements achieved with our approach. Later on we use data from the Sloan Digital Sky Survey, which has to do with the cosmological redshift of galaxies, as a real life example to further prove results from previous tests and to show the limits of median based sparse grids.