On Tue, 2007-07-24 at 16:59 -0400, Shiyi Chen wrote:
We are considering to use scaLAPACK routine PDPOTRF for Cholesky
factorization of dense N by N positive real symmetric matrix. Could you give
us a rough idea on the maximum size of matrix the software allows? We have a
cluster computer of 128 nodes and each node has 2GB CPU memory.
Alonzo G. Decker Jr. Chair in Engineering and Science
Department of Mechanical Engineering
The Johns Hopkins University
124 Latrobe Hall
3400 N. Charles Street
Baltimore, MD 21218
the Cholesky factorization in ScaLAPACK is done by means of the P_POTRF
(the underscore replaces the letter that defines the precision)
subroutine. The only area that needs to be allocated for this operation
is the memory space for storing the matrix. No other significant memory
is allocated internally. Thus, for a matrix of size N
NxN*4 bytes in signle precision
NxN*8 bytes in double precision
Note that in the cases where the matrix is banded or tridiagonal a lot
of space can be saved. In these cases the subroutine to use are P_PBTRF
and P_PTTRF respectively.
This said, it is very difficult to say how big of a matrix you can tun
on your system because this may depend on many other things like how
much space is you operating system using etc.
Since the processor count is relatively low and the Cholesky
factorization is a cheap operation, I suggest you to find the maximum
size by inspection. My guess is that you should be able to use at least
a matrix of size 100000 in double precision which makes a local size of
625 MB more or less.
Alfredo Buttari PhD,
Innovative Computing Laboratory,
UTK Computer Science Dept.
1122 Volunteer Blvd. Knoxville, TN 37996