Long integers in GSMap needed for large grids

[rljacob]

(copied from MCT's old bugzilla)

User's with large grids need long int support in GSMap.
"I am using MCT to perform communication and remapping operations between two
dummy models. For that tests, I'm defining the domain sizes, let say,
32768x32768 and 49152x49152. That produces a sparse matrix for the
interpolation with size in the order of 1.07 billion and 2.4 billion of points
respectively. However, when I compile MCT in a Cray XE6 machine (Hopper) and I
try to run it, I got negative values for the number of points indicating the
integer data type is not long enough."

"What I'm doing is very simple. A 2048x1024x1024 3D array is redistributed from
M processes of component 1 to N processes of component 2. In this case the
total number of grid points is 2048_1024_1024=2147483648, which overflow the
32bit integer. As a result, the init of MCT global map gave the following
errors:
"m_GlobalSegMap::initp_: non-positive value of gsize error, stat =-2147483648""

[jedwards4b]

The mct_gsmap_orderedpoints rountine will return an integer array with overflowed values without reporting an error.

[rljacob]

This sounds like a bug. How did you trigger it?

[jedwards4b]

Found it in a call from seq_io_mod - before I go about trying to reproduce
it I'm pretty sure that you will
see the problem by just looking at the code in OrderedPoints in
m_GlobalSegMap.F90

On Fri, Jul 10, 2015 at 1:24 PM, Robert Jacob [email protected]
wrote:

This sounds like a bug. How did you trigger it?

—
Reply to this email directly or view it on GitHub
#31 (comment).

Jim Edwards

CESM Software Engineer
National Center for Atmospheric Research
Boulder, CO