[Pw_forum] Too many data communications between nodes when run pw.x in parallel?
akohlmey at cmm.chem.upenn.edu
Mon Dec 29 19:21:00 CET 2008
On Tue, 30 Dec 2008, Wang Qinjing wrote:
QW> Dear all:
QW> I noted that if I made calculation by pw.x in paralllel, there were
QW> too many data communications between nodes. That was attributed to
QW> the data communication of wave functions. There is a severe delay to
QW> get information from these nodes.
please provide more detail about the kind of job you
are running, how you were running it and the kind of
hardware you are running it on.
there are several ways how you can improve parallel efficiency.
please see the documentation and the many previous discussions
on this topic and then make some scaling and performance tests
to see how to run most efficiently. the best settings can vary
a lot between different inputs and calculation types.
QW> If I set disk_io='low', things are much better. But the main node
this is only one option and indicates that you are most likely
not running your job in a very efficient manner.
QW> (or the first node) still has problem in net communication. I
QW> noticed a charge density file is still written after each scf step.
QW> Is that the problem? Or is there anything I can do to improve this
there is a lot that can be done, but since you don't provide any
information on what you are already doing, it is impossible to
give any recommendation beyond urging you to (re-)read the
documentation and FAQs.
QW> Thanks in advance.
QW> Yours Wang
QW> Wang Qinjing, Physics Department,
QW> Fudan University
Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu http://www.cmm.upenn.edu
Center for Molecular Modeling -- University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
If you make something idiot-proof, the universe creates a better idiot.
More information about the users