Prof. Dr. Gerhard Zumbusch
Explicit Finite Differences schemes for wave equations
and the return of the vector processor
Time: 17th July, 2013 at 11:00 h
Place: HS 11.02/ground floor
Institute of Mathematics and Scientific Computing
Heinrichstraße 36, 8010 Graz
Abstract:
Explicit Finite Differences schemes for wave equations and the return of the vector processor.
many numerical relativity code and wave simulation codes in general are based on explicit
finite differencing. The computational efficiency of such parallel multi-patch codes heavily
depend on the single node performance. Recent computer architectures include numerical
accelerators, based on graphical processing units (GPU) and parallel vector processors (Intel
Phi, Cell). Along with current microprocessors (Intel Sandy Bridge and AMD Bulldozer) they
rely on data and instruction parallelism on man levels. Together with the introduction of
accelerators came novel parallel programming paradigms (Cuda, OpenCL, OpenAcc, OpenMP
with offloading). It is by no means obvious how a portable and /or efficient finite difference
may look like. However, memory and cache aware algorithms, adapted to vector processors
show to be fundamental.