[Pw_forum] -D__NEW_PUNCH option
konstantin_kudin at yahoo.com
Thu Dec 2 18:14:08 CET 2004
> Of course, the local scratch
> disks are preferable. But after that when the code will finish, the
> files in the scratch directory will be deleted, and this is not
> preferable for me because I may have to use these files for restart.
I looked into the logistics of such a thing some time ago. We have a
cluster where each node has a scratch space uniformly named as, say,
/tmp. These directories were also accessible from any node by a longer
Basically, the wavefunction files split across nodes (*.wfc) in each
/tmp are not worth bothering with, so just discard them. They are easy
to regenerate if *.rho and *.pot are available. Other useful files only
exist on the main node for a given run.
Immediately before the job is run a wrapper script copies files with a
given prefix from a central location to the main node /tmp so that the
restart information can be read. After the job is finished, the same
wrapper script copies back files with a given prefix from the main node
/tmp to the central location. All the *.wfc files are discarded at this
stage. I could provide my setup for such manipulations.
Alternatively, perhaps, one could modify PWSCF to have 2 locations for
temporary files. That should be an easy hack. One location would have
everything but the *.wfc files, the other location would be just for
*.wfc files. Then these *.wfc files could go to the local scratch
space, and discarded whenever. The rest of the files could even go to a
central location. I think that using NFS for files other than *.wfc
would be reasonably mild for the network load. You could try this and
see if you admins are still unhappy.
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
More information about the users