High Performance Computing II
- Contents: Fast parallel
solvers for large (linear) systems of equations are investigated
by the student.
- direct solvers and approximative inverse
- iterative solvers and preconditioners
- presentation of the students on the investigated solvers
- comparison of the solvers above on various platforms as cluster
computers, shared memory computers,
distributed memory computers
- Final presentation and Tech-report of the results
- Lecturer:
- Prof.
Gundolf
Haase, Heinrichstr. 36, Zi 506, Tel. 5178,
- Lecture Time:
- Thursday 10:15 -
11:45 in SR 11.32
- Computers:
List
- Solver and other packages
(a selection from Arnold Neumaier's page,
some literature)
-
- partitioners: METIS,
Chaco,
JOSTLE,
MESHPART
(Matlab Mesh Partitioning and Graph Separator Toolbox,
with Matlab interfaces to METIS and Chaco,) PT-Scotch
- direct solvers: MUMPS,
SuperLU, PaStiX,
CHOLMOD
[new]
- preconditioners: Hypre,
AMG by Peter
Arbenz;
approx. inverse: pARMS,
SPAI,
ParaSails, AHMED,
HLib
- iterative solvers: PETSc,
Aztec,
- load balancing, parallel data management: Dune,
Zoltan
- Projects/Talks (SS12):
- 29.03.2012: Presentation
wrt. Intro into mesh partioner PT-Scotch,
overview on the Trilinos
package
- 21.06.2012: Detailed Presentation
on Trilinos and PT-Scotch
(Jonathan
Hu wrt. Trilinos: Tpetra provides sparse linear algebra
operations between compute nodes. Kokkos provides specific kernels
that
can be optimized to run within a compute node (GPU or multicore))
- Data:
- 2 example matrices and read
routine for it (129 MByte, thanks to Aurel
Neic)
- 2 examples meshes (thanks to Aurel
Neic):
- Compute-Server:
-
- MATH047016.uni-graz.at
(dualhex: 12 cores: 2 Opteron 2427; 2.2 GHz; 28 GB memory),
Ubuntu 11.10
- MATH047201.uni-graz.at
(quadquad: 16 cores: 4 Opteron 8347, 1.9GHz; 32 GB memory): 201
- mephisto
(math047128)
- last update: June 28, 2012