[Mitgcm-support] configuring MITgcm for atmospheric work
mitgcm-support at dev.mitgcm.org
mitgcm-support at dev.mitgcm.org
Wed Jul 9 15:47:07 EDT 2003
Hi Amy,
I've been very busy for the last three weeks. Hows it going?
Are you getting anywhere? Heres some "answers" to the
questions you sent in October.
Q1 - I think a 4-way SV1 could compete with an 8 or possibly
16 way O2000. However, we don't do a lot of work on the
SV1 (there are only a few built) so I only know
the marketing numbers.
Q2 - The usingMPI parameter is set true whenever ALLOW and
ALWAYS USE_MPI are defined. Thats just the way it is.
Q3 - It probably is possible to get C90 performance at T21 on
an SGI at around 8 procs. Not sure why you get a discrepancy.
If you send the timing info that is written at the end
of the STDOUT files that might have some clues. What sort
of timings do you need to get your work done?
Q4 - There is still a lot of research ongoing on the cube sphere
grid. Getting something up on latlon first is probably
best. Switching later is always possible.
How are the physical solutions looking?
Chris
> amy wrote:
>
> Hi Chris,
>
> Thank you for answering my email so quickly. I would like to get the
> model up and running so I can find out if I would like to use it for
> my
> study of decadal variability in the Pacific basin.
>
> I have an SV1 and an Origin 2000 that I can run the model on.
> Q1:Do you think that running the model with 8-16 cpu on the SGI
> would be faster than running the model with 1-4 cpu on the SV1?
> Q2:When do you set the usingMPI parameter to true?
>
>
> I ran the aim_5l.LatLon configuration using a T21 grid yesterday and
> if
> I increase the time step to 900 seconds I can ran 1 model day in 10
> seconds.
> Q3:Should I be using the cubic sphere configuration instead?
> Q4:The webpage which describes the atmospheric physics says that
> their T21 run on the C90 took 1 second to run 1 model day. Shouldn't I
> be getting similar performance with this setup?
>
> Thank you for help!
> Amy
>
More information about the MITgcm-support
mailing list