Page 1 of 1
general matrix parallel eigensolver

Posted:
Fri Jun 13, 2008 2:51 pm
by mkrems
there is no routine for finding the eigenvalues and eigenvectors for a general real matrix in scalapack....how is this done?

Posted:
Fri Jun 13, 2008 3:05 pm
by Julien Langou
There is no easy way to have a full nonsymmetric eigenvalue solver from
the ScaLAPACK software. Julien.

Posted:
Fri Jun 13, 2008 3:21 pm
by mkrems
thanks for the quick reply. what is the hard way then? is this just something people don't do in parallel?
Re: general matrix parallel eigensolver

Posted:
Thu Jul 17, 2008 7:34 am
by hurlements
In fact, the answer could be interesting for me as well. The codes I am currently writing rely
on the resolution of an eigenvalue problem for a general complex matrix. I am wondering
if there exist any parallel solver for such problems.
If you know anything about it, please, let us know. There are plenty of things that
can be done, on the web - but very scarcely things that are not possible.
Yours,
Antoine.
Re: general matrix parallel eigensolver

Posted:
Thu Jul 17, 2008 8:11 am
by hurlements
I think I have found the answer : nothing is ready yet for non symmetric eigenproblems
in scalapack, but this is likely to come (I found a power point presentation in which
the summer 2007 was mentionned to begin with that...).
http://www.cs.berkeley.edu/~demmel/futu ... PARA06.pptSince everybody uses scalapack, nobody seems to have any parallel solver for
such problems. But one day...
Antoine.
Re: general matrix parallel eigensolver

Posted:
Thu Jul 17, 2008 8:55 am
by hurlements
If I have understood what I've just read :
There is no equivalent of zgetri nor zgeev with scalapack because these problems
do not scale well - or this is not done. But there is still a way to take advantage of
a multi-proc computer by using a multithreaded blas library - which will already
greatly improve the speed of the code.
Suggestion : atlas or goto blas, that's it ?
This is not obvious which BLAS are multithreaded and which are not...
Antoine.