[Pw_forum] how to control the writing of wavefunction
Lorenzo Paulatto
paulatto at sissa.it
Thu Dec 13 14:16:01 CET 2007
On Thu, December 13, 2007 13:42, xu yuehua wrote:
> yes,maybe once i set disk_io="none" ,when something goes wrong ,like
> power
> is off suddenly,you have to start from scratch .from this point ,it is
> wasting time in deed.
> but is there an better idea ?
> i am wondering
I think it is highly unlikely that you have more RAM than disk space, so
disk space should be no issue. Furthermore it should not be so slow, as
long as you are leaving wfc_collect=.false. which is the default.
>From these assumptions I can guess that you are using a network share as
temporary directory for your calculation; in other words your outdir is
set to some place, probably in you home directory, which is not local to
nodes but stays on a server and is shared via NFS (network file system).
If this is your case, every time wavefunctions are written the data has to
pass trough the network, it is somehow equivalent to collect_wfc=.true.
but done externally by the operating system. Furthermore if you are using
the same network for NFS and mpi when the traffic from NFS suddenly bursts
it is likely to conflict with mpi and slow it down a lot.
To avoid this problem you must use a local directory as outdir (like
/tmp), but you to have to remember that if you want to restart a job you
have to use exactly the same node you used before in order for them to
find their wavefunctions. Otherwise you may have to recollect the data by
hand, which is probably not trivial (I have never tried).
I hope what I wrote was not already known to you, goodbye.
--
Lorenzo Paulatto
+39 040 3787 511
http://people.sissa.it/~paulatto/
----------------------------------------------------------------
SISSA Webmail https://webmail.sissa.it/
Powered by SquirrelMail http://www.squirrelmail.org/
More information about the users
mailing list