A question about ScaLAPACK

[fengsh]

I'm now using LAPACK to solve a huge non-hermitian complex eigen matrix.
It works perfectly. I would like to know if I could use ScaLAPACK to
solve the same problem with parallelization?

If yes, do you have any idea how much efficiency it will be? For
example, use 16 CPUs? Is there any rough number that I could estimate the improvement?

Many thanks,
Feng

[Julien Langou]

Hello,

Unfortunately there are no driver for the standard eigenvalue problem in ScaLAPACK.

There are some efforts in trying to provide one soon (soon ~ a year).

There are currently some partial routines that enables one to compute the solution of the
standard eigenvalue problem: (1) reduction to Hessenberg form (PZGEHRD) (2) solution of
the Hessenberg problem (PZLAHRD). Combining the two together should give you a
temporary solution, unfortunately this turns out to be farily hard.

Best wishes,
Julien.

[fengsh]

Hi Julien:
Thanks for your reply. By standard eigenvalue problem, you mean the complex non-hermitian eigen matrix, right?
Will PLAPACK be capable to solve this problem?

Best wishes,
Feng Shaoi

[Julien Langou]

Hello,

By standard eigenvalue problem, you mean the complex non-hermitian eigen matrix, right?

correct. (I meant Standard as opposed to Hermitian or Generalized or etc.)

I am pretty sure there is no true parallel distributed standard eigensolver existing.

As far as PLAPACK goes, I'd say no, the routine is not in there. Best is that you ask
the PLAPACK developers by yourself if you really want to be sure. If you do get an answer,
let us know what it is.

Julien.