VSR-Seminar

Vortrag 1: Lattice Boltzmann simulations of complex fluids

Speaker: Prof. Dr. Jens Harting, IEK-11

In this presentation I will summarize recent results from lattice Boltzmann simulations of complex fluids obtained at the Helmholtz Institute Erlangen Nürnberg (IEK-11). I will present the latest improvements and scaling data for our simulation package LB3D and then focus on large scale simulations of capillary suspensions, self-assembly of colloidal particles at liquid interfaces, the dynamics of evaporating droplets, and multiphase flows including electrokinetic effects. For the latter, we recently developed a fully parallelized solver for the dynamics of charges and charged colloidal particles which is coupled to the multiphase lattice Boltzmann solver.

Vortrag 2: Neuroimaging Analytics in large cohort datasets

Speakers: Dr. Jan Schreiber, Prof. Dr. Svenja Caspers, INM-1

In-vivo neuroimaging has experienced a paradigm change, from focused studies in a group of 20-30 subjects to now large population-based cohorts with several thousands of subjects, enabling investigation of various environmental and genetic influencing factors on brain structure, function and connectivity. For processing and analyzing these large amounts of data using multi-dimensional statistics, new computing solutions are needed, as present approaches on local computing clusters are not feasible anymore, with computing times up to several months for a single study.

The aim of the current compute time proposal is thus to setup efficient workflows and processing pipelines for subsequent analysis of neuroimaging data on structure, function, and connectivity of the human brain using established software tools. With one subject being the relevant entity for data processing on the first level, i.e. the preprocessing of all data of each subject, the major focus is here on optimizing workflows for doing the computations for many subjects in parallel. This poses a challenge for file management as hundreds of single files are produced, written and read constantly. Parts of these workflows are parallelized (using MPI and OpenMP), in particular when permutations or hundred thousands of repetitions are needed for generating probability distributions. This is also partially implemented on GPUs. Especially these latter approaches also apply to the group level statistics, where the preprocessed data from the thousands of single subjects are aggregated and analysed in relation to environmental and genetic factors.

Currently, most of the single-subject preprocessing workflows are running and data of two large cohorts (with more than 1,000 subjects each) are being processed. This will be continued with even larger cohorts (with 30,000 and 100,000 subjects, respectively) as well as followed by dozens of studies using the group level statistical approaches (where for each study, the group level computations have to be done anew, depending on the research question), for which future computation time on JURECA is envisaged.