[Pw_forum] k point parallel

[Konstantin Kudin]

> yes. please have a look at the manual. you need to give the
> additional
> flag -npool to the executables. this decides, how many k-point pools
> will be used. note that this should be a divisor of the number of
> cpus
> requested and the number of k-points. for optimal performance, you
> also need a _local_ scratch area on each node (and specify that
> in your input) and set the keyword wf_collect=.true. so that the
> final restart is collected to a shared directory, or else you would
> most likely not be able to restart or continue with the next job, as
> you would need to have the exact same nodes in the exact same order
> assigned for that job.

 Axel,

 Good tips! Also note that most PWSCF jobs can be restarted from 'rho'
files alone without 'wfc' files (with very minor extra cpu time), thus
saving the effort to collect and manage the wavefunction file.

 I would also like to add that the upcoming "official" version of
Espresso will allow to specify separate directories for 'wfc' files.
This way they can go to local nodes, while the other restart info goes
to some other directory perhaps even using NFS.

 Kostya


	
		
__________________________________ 
Yahoo! Mail - PC Magazine Editors' Choice 2005 
http://mail.yahoo.com