pnetcdf strange behavior

Christopher Subich csubich at math.uwaterloo.ca
Thu Apr 12 18:55:53 CDT 2007


Hongzhang Shan wrote:
> slower from 7GB to 8GB. Does this mean something goes wrong with my
> code or the first 7GB is only allocated without actually writing the
> data.

The file must be allocated up to the Nth bit before that bit can be 
written.  One (likely) possibility is that process 7 writes to the file 
fairly early, meaning that the file is (sparsely) allocated up through 
that process's contribution.  Whenever process 8 gets a turn, the 
remainder of the file is allocated and written.

You can see this working for yourself; test total writes (for 8 
processes) of 2GB/4GB/8GB, and you should see essentially linear scaling.




More information about the parallel-netcdf mailing list