Subdomain division

21 views
Skip to first unread message

fatemeh chegini

unread,
Jan 18, 2012, 5:13:06 AM1/18/12
to getm-...@googlegroups.com
Dear GETM users,

I'm trying to divide my domain using the subdomain utility. I'm running my model on a cluster in which the connection between the sills are slow and so it's important that the distribution of the divided domain be reasonable. Is there a way that I tell the subroutine to assign all the neighbor domains to one sill even if the load distribution is a bit messed up?  I've tried setting different kinds of parameters but it doesn't seem to get better.

Best Regards,
Fateme

Knut

unread,
Jan 18, 2012, 5:24:12 AM1/18/12
to getm-...@googlegroups.com
Dear fatemeh,

fast answer, but not related to the subdiv tool... :-)

With ifort I experienced that the automatic vectorisation (enabled by
default with -xHOST) for one process per node *can* be faster than
having several mpi processes on a single node (of course depending on
the water/land distribution in your domain). Maybe you want to try out
that as well ...

Cheers,
Knut

Ulf Gräwe

unread,
Jan 18, 2012, 5:45:57 AM1/18/12
to getm-...@googlegroups.com

You have to set: load_balance_method=2 !
This will tell subdiv to reduce inter node communication and will try to
put neighbouring subdomains on the same node. However, it still tries to
optimise for the total number of wet cells on each node.

Cheers, ulf

--
Ulf Gr�we

Leibniz Institute for Baltic Sea Research Warnemuende
Dept. for Physical Oceanography and Instrumentation
Seestrasse 15
D-18119 Rostock-Warnemuende
Germany

Pho. : +49 (381) 5197-358
Fax : +49 (381) 5197-440
Web : http://www.io-warnemuende.de/ulf-graewe-en.html
Room : 216

fatemeh chegini

unread,
Jan 18, 2012, 5:55:27 AM1/18/12
to GETM-users

Dear Ulf,

Thanks a lot.

>
> You have to set: load_balance_method=2 !
> This will tell subdiv to reduce inter node communication and will try to
> put neighbouring subdomains on the same node. However, it still tries to
> optimise for the total number of wet cells on each node.
>
> Cheers,  ulf

Dear Knut,

Thanks, I'll try that.

> With ifort I experienced that the automatic vectorisation (enabled by
> default with -xHOST) for one process per node *can* be faster than
> having several mpi processes on a single node (of course depending on
> the water/land distribution in your domain). Maybe you want to try out
> that as well ...
>
> Cheers,
> Knut

Best,
Fateme
Reply all
Reply to author
Forward
0 new messages