Hello,
I have a question concerning the matrix descriptors in SCALAPACK.
I'm using the SCALAPACK function PZGELS to solve a complex linear equation
system in a distributed memory system.
As I'm doing the administration of the needed memory myself, I calculate the
exact sizes of the local matrices.
To illustrate that, I give you the following example:
9 x 9 matrix with a 2*2 blocking and a 2*3 process grid like in the example 1,
SCALAPACK user's guide.
Then the sizes of the submatrices are as follows:
process row x process column |size |local leading dimension
0 0 |5x4 |5
0 1 |5x3 |5
.
.
1 2 |4x2 |4
(everything equivalent to the cited example from the user's guide)
I fill the matrix descriptor of A, here especially the maximal local leading
dimension of A (entry number 9) with the so calculated exact values.
But this always leads to a problem when I try to use the function PZGELS.
When I set this value (maximal local leading dimension of A) in each
descriptor to 5 (or to any higher value) the calculation starts.
First I thought this may be a problem of an insufficient size of this
parameter, but each exact value multiplied by 10 per example doesn't work
neither.
So obviously the descriptor (at least entry 9) has to be EQUAL in each local
sub matrix.
Is this right? Are their possibilities to use the exact values?
My matrices are quite big and it would be good to use local sizes to allocate
the local matrices and all that.
I would be very grateful for any help or any suggestions you may have,
With best regards,
Jana Hutter
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.eecs.utk.edu/mailman/private/scalapack/attachments/20100728/65261308/attachment.html
|