[Mitgcm-support] RE: Global Files / MIT Model

[mitgcm-support at dev.mitgcm.org]

Hi John,

  Sounds like an OK fix to me. Would you be happy to
share/donate your netcdf code? If we are going to revisit
mdsio I would like to at least think about introducing the option
to put out netCDF directly. If you have something that
is going in that direction, that would be very useful starting
point for us.

Thanks,

Chris
-----Original Message-----
From: adcroft at ayana.lcs.mit.edu [mailto:adcroft at ayana.lcs.mit.edu]On
Behalf Of Alistair Adcroft
Sent: Wednesday, January 23, 2002 9:30 AM
To: John D Stark
Cc: support at mitgcm.org
Subject: Re: Global Files / MIT Model


Hi John,

mitgcm.org had a disk problem last week but it back up now. I'll CC this
there.

I agree with your assessment of the problem and we'll make those changes
fairly
soon.

Thanks,

A.

John D Stark wrote:
> Hi Alistair,
> I sent the following to support at mitgcm.org, but it bounced ('Connection
refused
> by mitgcm.org')  so, I'm sending it directly to you.  Is there a new
support
> news group (it's no longer visible)? Thanks for the information you gave
me
> earlier this month.
>
> Cheers,
> John
>
> Date: Fri, 11 Jan 2002 18:02:43 +0000 (GMT)
> From: John D Stark <jods at socnet.soc.soton.ac.uk>
> Subject: Reading / Writing Global files
> To: support at mitgcm.org
> MIME-Version: 1.0
> Content-MD5: CySUjCN5SYzpljLstJ0M6Q==
>
> Hi,
> I spoke to Alistair earlier this week, who thought you may be interested
in some
> of the coding I have been doing here in Southampton.  I hope this is
useful for
> you.
>
> Although global files are somewhat slower to write they are much easier to
deal
> with, so I use them a lot.  However, with threads or MPI they just don't
work.
> The main problem is that writing by multiple threads/ process to a single
file
> is not safe, and the mdsio routines don't control this.
>
> What is needed with global files is a large loop over processes within
> mdsio_writefield with a (MPI_)barrier at the bottom.  The loop / barrier
should
> be dependent on whether globalfiles is true, which would minimise loss of
speed
> with multifile writing.
>
> I found that putting this in corrected problems I had with missing data in
> checkpoint files, and was also required for some netcdf file i/o I have
added.
>
> Note that adding a barrier in mdsio would mean that all higher level
routines
> that enclose I/O in a begin_master / end_master section will cause a
problem,
> notably write_state.
>
> Cheers,
> John
>
> ------------------------------------------
> John Stark
> Applications Programmer
> Southampton Oceanography Centre
> Tel. +44 (0)23 8059 6571
> e-mail jods at soc.soton.ac.uk
> ------------------------------------------