Abstract (EN):
The advantages of using parallel libraries to support the programming of mathematical models in High Performance Computing is illustrated in this paper. We will briefly describe a model for a radiative transfer problem that involves an integral operator on a very large interval. Its discretization leads to very large matrix problems and, therefore, the use of parallel distributed computing is crucial. In order deal with the large dimension of the problem, we use an iterative refinement strategy to improve the eigenvalues and vectors of a moderate size matrix discretization computed as initial approximations, and thus provide accurate eigenpairs corresponding to a larger dimensional problem. Both stages require many matrix-vector multiplications with matrices that are built and stored in a distributed mariner among the processors, due to their large size. Computational details of the use of PETSc and SLEPc packages on the solution of this spectral problem will be addressed.
Language:
English
Type (Professor's evaluation):
Scientific
Contact:
pjv@fe.up.pt; jroman@itaca.upv.es; falmeida@fe.up.pt
No. of pages:
3