Generalized Eigenvalue Problem for Very Large Matrices

Method:

• very large sparse matrices from finite element computations
• condensation:
  • find most important degrees of freedom(``masters'')
  • reduce problem to masters, capturing main aspects of slaves
• involved hierarchical data structures
• reduction of linear dimension typically by a factor of $10^2$
• condensed problem is dense

Parallelisation:

• master/slave program for the condensation of partial structures
• uses BLACS and basic MPI for communication
• Scalapack for the condensed generalized eigenvalue problem