The “Seminar on efficient numerical simulation on multi- and manycore processors” (MuCoSim) covers topics from optimization, parallelization and analysis for modern multi- and manycore systems. The assignments are chosen from interesting contemporary problems in High Performance Computing. Experiments are conducted on a range of modern hardware architectures, from multicore processors over GPUs to vector machines, and also large super computers. Deliverables are two talks and a written seminar report.
This is a (non-exhaustive) list of possible topics for the MuCoSim seminar.
- We have a NEC Aurora “Tsubasa” vector system installed at RRZE. This is a true vector computer architecture, and we are working closely with NEC to get to know the machine and explore its potential for scientific computing.
- Porting the Mantevo Conjugate Gradient (HPCCG) benchmark to Tsubasa and analyzing its performance properties
- Using the “Stempel” automatic stencil code generator tool to analyze stencil algorithm performance on Tsubasa
- Analyzing sparse matrix-vector multiplication (SpMV) with the SELL-C-sigma format on Tsubasa
- Using OpenMP 4.5 accelerator directives to port the Mantevo HPCCG benchmark to GPGPUs and analyze its performance properties
- Distributed-memory parallelization and performance analysis of SpMV with MPI on multicore clusters
- Parallelizing and analyzing a Gauss-Seidel algorithm with OpenMP and tasking
- Writing a simple OpenMP performance analysis tool using the OMPT interface
… and more.