[Wien] MPI Strategy

[Peter Blaha]

Please read the doku about parallelization in the WIEN2k usersguide.
mpi-parallelization is NOT done over k-points.

mpi-parallelization is useful ONLY for bigger cases, where big depends
on your machine, accuracy and number of processes you want to use.
Typically, the blocksize of an mpi-block should still be at least 1500x1500,
better a bit larger.
If your matrix size is 1, you may use up to 6x6=36 (or 6x4=24 in your case) 
cores, which gives you
1/6 as blocksize. Probably scaling becomes already problematic ...

If your matrix size is smaller/bigger, you have to adjust the number of cores 
which are
useful. Eventually, mpi-parallelization is NOT useful at all.


When you say: This machine does not allow passwordless ssh

what exactly do you mean by that ??

I guess you have to use a queuing system ?? Which one ??
Could it be, that ssh is blocked interactively, but it works fine within a
batch-job ?
Sometimes, rsh works within batch-jobs,...
You need to have some doku or some expert on the specific machine.
I'm pretty confident that there are rather simple ways to overcome any
restrictions due to a special sysadmin-setup, but most one needs to
know all kinds of details and testing ...

Am 01.04.2011 14:05, schrieb David Tompsett:
 Dear All,

 I have recently moved my Wien2k calculations to a new machine (Hector in the 
 UK). This machine does not allow passwordless ssh and therefore disables the 
 k-point parallel part of
 Wien2k. I therefore want to clarify some points about how MPI (chiefly 
 lapw1_mpi) works in Wien2k.

 If I run a calculation with say 5 k-points using 5 nodes that each have 24 
 cores, then will the MPI call of lapw1_mpi split the 5 k-points over the 5 
 nodes? Or will it simply run
 the k-points one at a time, using all 120 cores for each one?

 Is there any documentation I can read about the MPI parallel strategy more 
 generally?

 Many thanks,
 David Tompsett.



 ___
 Wien mailing list
 Wien at zeus.theochem.tuwien.ac.at
 http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien

-- 

   P.Blaha
--
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671 FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.atWWW: http://info.tuwien.ac.at/theochem/
--

[Wien] MPI Strategy

[Laurence Marks]

I've seen this sort of trend elsewhere, sysadmins expecting users to only
run one mpi job rather than parallel ones, and this can lead to problems for
Wien2k. Some possibilities:

1. Run everything as 1 mpi. This is less efficient, quicker for you.
2. Use rsh, or ask a sysadmin for an equivalent.
3. Set mpi_remote and/use_remote to 0. The different mpi will then be
launched from one node but should execute according to your machines
breakdown.
4. If desperate, and comfortable with shell programming, edit parallel
options so it in fact runs a csh or sh on remote nodes. This is needed for
some openmpi because it does not export some environmental variables.
On Apr 1, 2011 7:06 AM, David Tompsett dat36 at cam.ac.uk wrote:
-- next part --
An HTML attachment was scrubbed...
URL: 
http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20110401/cd90a30c/attachment.htm

[Wien] MPI Strategy

[David Tompsett]

Dear All,

I have recently moved my Wien2k calculations to a new machine (Hector in the
UK). This machine does not allow passwordless ssh and therefore disables the
k-point parallel part of Wien2k. I therefore want to clarify some points
about how MPI (chiefly lapw1_mpi) works in Wien2k.

If I run a calculation with say 5 k-points using 5 nodes that each have 24
cores, then will the MPI call of lapw1_mpi split the 5 k-points over the 5
nodes? Or will it simply run the k-points one at a time, using all 120 cores
for each one?

Is there any documentation I can read about the MPI parallel strategy more
generally?

Many thanks,
David Tompsett.
-- next part --
An HTML attachment was scrubbed...
URL: 
http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20110401/727edd5b/attachment.htm