pnetcdf question

[Rob Ross]

James,

Bill's note here is right on the mark, and this is actually very similar 
to the issue that Roger Ting brought up just a little while ago.  You're 
much better off doing a little communication to get the offsets right and 
then performing a collective I/O operation.

Regards,

Rob

On Thu, 18 Mar 2004, William Gropp wrote:

> At 01:59 PM 3/18/2004, Foucar, James G wrote:
> >Hi Rob, thanks for the reply. i was thinking along the same lines. the
> >reason i am hestitant to do the:
> >
> >array 1: 1a 1b 1c 2a 3a 3b 3c 3d 4a 4b 4c
> >array 2: 3 1 4 3
> >
> >implementation is because it will require communication between the
> >processors to figure out what offset they should be writing to. so, right
> >now i am leaning towards separate variables per process. this will require
> >me to enter "independent data mode", correct? will this cause me to take a
> >performance hit?
> 
> If all you need to do is to compute the offset, the  MPI_Scan routine will 
> do this fairly efficiently (if you have a good implementation).  If you 
> have an MPI-2 implementation, MPI_Exscan is a closer match to what you 
> need.  The communication cost should be small relative to the I/O cost.
> 
> Bill 
> 
>