[Mitgcm-support] Re: TAF
mitgcm-support at dev.mitgcm.org
mitgcm-support at dev.mitgcm.org
Wed Jul 9 15:53:45 EDT 2003
Dear ECCO & Co.,
here are several comments regarding recent e-mail exchanges,
(cf. appended mail plus one that I'll post separately),
combined with my apology if my latest updates weren't clear
enough and didn't include some people,
and with a suggested New Year's resolution to revive
communication amongst ourselves when there are problems.
1.
There seems to be a misunderstanding.
Currently we at MIT have the same version as SIO (1.4.17).
It only became clear this week (Dec. 31st) that this version
creates erroneous adjoint code for MITgcm with GM, and that it's
indeed TAF which is responsible for the erroneous adjoint GM code
(as of Dec. 21st it was agreed between Ralf and me that there is
a bug in the adjoint code when compared to tangent linear and
finite difference gradients, but it wasn't clear where that bug
came from).
2.
I understand that version 1.4.19 which Ralf referred to was
his development version which was not yet publicly released,
and that 1.4.20 which Ralf referred to does not yet exist,
but will be the new version with Ralf's latest fixes,
and will be implemented ECCO-wide as soon as it's available
and tested.
3.
Ralf has agreed to add this version/setup (MITgcm c47 with GMRedi)
as a quality control or additional benchmark when upgrading TAF.
We should keep this process of well defined benchmarks,
(adjoint verification experiments) and "upgrade" those with
the addition of new/modified/problematic code and setup.
4.
In the previous context it is news to me that there were problems
with KPP when using TAF, but none when using TAMC.
Here, a well-defined test setup (finite-difference/tangent linear
gradient checks) which exhibits the problem would have been helpful.
5.
It would be useful for us to get more details on what went wrong
with TAF (to give us a better understanding of its working
and where to suspect bugs).
Also, a more similar code output between TAMC and TAF
would help tracing problems in either adjoint code.
6.
We should think again about reviving an ECCO email list.
We had implemented and advertised two mailing lists
(see the ECCO meeting on May 31st/June 1st 2001)
ecco-devel at mitgcm.org (for model developers/users)
ecco-all at mitgcm.org (for official announcements)
which didn't seem to appeal.
Therefore, all recent communication regarding MITgcm
and adjoint development/issues went to support at mitgcm.org
(all mails are visible through the corresponding newsgroup).
I stop here before it gets too lenghty.
Cheers
-Patrick
Detlef B. Stammer wrote:
> Ralf;
>
> from your exchange with Patrick I see that the TAF version 1.4.17 creates
> erroneous adjoint code and that you have later corrected versions. I also see
> that we at SIO never go the later versions installed despite the fact that you
> seem to know that we have incorrect adjoint codes generated by TAF. Neither did
> we get EVER a warning from you.
>
> I assume that a correctly run company needs to send out warning messages
> immediately to users and that under our license contract you are required to
> update our TAF code as soon as you have new - in fact correct - versions
> available.
>
> The latest update we had is from Nov. 7 2002.
>
> Is there any misunderstanding on my side? Or can I assume that we will get
> version 1.4.20 installed in the very near future?
>
> Thanks a lot and Happy New Year!
> Detlef.
--
_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/
Patrick Heimbach Massachusetts Institute of Technology
FON: +1/617/253-5259 EAPS, Room 54-1518
FAX: +1/617/253-4464 77 Massachusetts Avenue
mailto:heimbach at mit.edu Cambridge MA 02139
http://www.mit.edu/~heimbach/ U.S.A.
More information about the MITgcm-support
mailing list