[Pw_forum] Fwd: help
giannozz at nest.sns.it
Mon Jun 12 09:43:18 CEST 2006
> From: Hüseyin Murat TÜTÜNCÜ <tutuncu at sakarya.edu.tr>
> In our computer system, different nodes have their own local disks
> (with scratch partition), and a rather slow communication link between
> master and slave nodes. The present version of the pwscf code assigns
> a single scratch file space (presumably at the master node). This
> really slows my calculations, as well as virtually stops other users
> from using the machine.
> I would be grateful if you could help me (i) identify scratch and
> re-start data files, and (ii) advise me how to utilize scratch disks
> on local (i.e. slave) nodes instead of using one scratch file space on
> the master node.
in parallel execution, each processor writes its own chunk of
wavefunctions, unless variable 'wf_collect' is specified in input.
Wavefunctions are usually by far the largest file. All the rest
(small files, except the charge density that becomes large only
for very large systems) is collected on a single processor and
written there. If this slows down the calculation, either you are
writing through the network (check where the directory pointed
by variable "outdir" resides: I/O through NFS must be avoided)
or you have really slow communications or disks or both.
The above holds for the current version of Pwscf (3.1)
Paolo Giannozzi e-mail: giannozz at nest.sns.it
Scuola Normale Superiore Phone: +39/050-509876, Fax:-563513
Piazza dei Cavalieri 7 I-56126 Pisa, Italy
More information about the users