[MITgcm-devel] Re: [MITgcm-support] noise in high resolution run
Martin Losch
mlosch at awi-bremerhaven.de
Sat Apr 1 11:10:53 EST 2006
Hi Chris,
the beers were good, the ambience even better (with Patrick in
Hamburg's smoke stacked basement bars that are decorated like your
grandparents living rooms in the fifties) and we tried to avoid
talking about noise in the MITgcm (did not quite succeed).
In my home directory on faulks there is a gzipped archive
~mlosch/dp.tgz [58MB]
of my configuration, it unpacks into a directoy drakePassage with a
"code" directory (please disregard btmvisc.tar: those are Matt's
modif's for increased viscosity near the bottom, which I don't use,
got in there by accident), "input" directory with initial conditions
and open and surface boundary conditions etc, and a "run00" directory
with my most recent data* files (most of them are just copies of the
versions in the input directory. The MITgcm code that I ran this with
was up-to-date on Thurday, Mar30, but I don't think that JMCs recent
changes will change anything (I'll update and rerun on Monday
anyway), so you can use up-to-date code for this, I guess.
EXCEPT: I defined USE_ISOTROPIC_SCALING in GAD_OPTIONS.h, so that
xd1-420-6::MITgcm> cvs -q diff
Index: pkg/generic_advdiff/GAD_OPTIONS.h
===================================================================
RCS file: /u/gcmpack/MITgcm/pkg/generic_advdiff/GAD_OPTIONS.h,v
retrieving revision 1.5
diff -r1.5 GAD_OPTIONS.h
31c31
< #undef ISOTROPIC_COS_SCALING
---
> #define ISOTROPIC_COS_SCALING
The noise only starts to appear after few (20) days of integration,
which take a while.
Good luck
Martin
On Mar 31, 2006, at 8:23 PM, chris hill wrote:
> Hi Martin/Patrick,
>
> Is it possible to transfer everything to a directory on faulks, we
> can take a more detailed look while you all drink beer.
>
> Thanks,
>
> Chris
> Patrick Heimbach wrote:
>> Hi all,
>> I am moving this to the devel list.
>> Given that we've been seeing also somewhat strange behaviour
>> in vertical velocity fields near topography and not so stable
>> behaviour of the adjoint as a consequence, I am wondering
>> (as is Martin) whether this is indicative of a bug.
>> Haven't looked carefully recently (in particular after the adjoint
>> bug fix for variable aliasing), will dot it soon, but don't think,
>> the 4deg adjoint showed this behaviour in much earlier versions.
>> The adjoint might actually tell us (if we believe it) where to
>> look for potential sources of the problem. I'll try to have
>> a look next week.
>> But maybe it's all ok...
>> OK, after making such provocative statement, I'll now
>> be offline for a few days (and have beers with Martin instead).
>> -p.
>> Quoting Baylor Fox-Kemper <baylor at MIT.EDU>:
>>> Hi Martin,
>>>
>>>> useJamartWetPoints is a good suggestion. I will try that next.
>>>> Since the noise seems to be associated with the coasts it
>>>> sounds like a good guess. What about the other flags,
>>>
>>>
>>>> useJamartMomAdv
>>>
>>>
>>> I was told that this one doesn't really make sense in terms of
>>> conservation near boundaries, I believe. I'm no expert in it,
>>> though.
>>>
>>>> SadournyCoriolis
>>>
>>>
>>> JMC just told me that this is actually NOT the scheme that
>>> Sadourny suggests to use, it is really one of the ones that he
>>> suggests NOT to use in the same paper! So, I think that for the
>>> moment this is probably not a good bet either.
>>>
>>>> I use
>>>> no_slip_sides=.false.,
>>>> the sidedrag code is a drag to debug. Each time I have a look
>>>> at it, it looks wrong the first time around and then after 3hs
>>>> of mind warp it turns out to be right. When I have a look at
>>>> the plots I find it hard to believe that there is not a bug in
>>>> the whatever- part of the code (doesn't even have to be the
>>>> viscosity). I guess I have to start turning off terms and see
>>>> what happens.
>>>
>>> sidedrag can't be the problem then...
>>>
>>> What I fear is what Michael suggests:
>>>
>>>> we had some problems like this with the NCOM (hydrostatic) at
>>>> one time and tried all sorts of things.
>>>> it turned out to be an indexing problem in the code.
>>>
>>> The question is, where?
>>> -Baylor
>>>
>>>> Dimitris,
>>>> I don't use KPP so far, although I would like to use in the
>>>> future (maybe, I have enough problems as it is (o:). My
>>>> experience with KPP is that it tends to amplify noise, but does
>>>> not generate it (?). In that case the horizontal filters are
>>>> useful. In this case, I don't think that it will help too much,
>>>> I am afraid, because the noise is already there without KPP.
>>>>
>>>> Thanks for your suggestions.
>>>> Martin
>>>> On Mar 30, 2006, at 8:06 PM, Baylor Fox-Kemper wrote:
>>>>
>>>>> Hi Martin,
>>>>> I've seen similar noise in W in some runs, but not so
>>>>> obviously in U and V. Have you tried jamartwetpoints?
>>>>> Also, are you using no-slip? We found a few bugs in the
>>>>> sidedrag code a while back. It might be worth revisiting...I
>>>>> suppose it is possible that there is a problem elsewhere in
>>>>> the viscosity/viscous terms code, but I don't know why it
>>>>> would be localized in space.
>>>>> -Baylor
>>>>>
>>>>>
>>>>> On Mar 30, 2006, at 8:30 AM, Martin Losch wrote:
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> I seem to have a problem with a 1/6 by 1/6*cos(phi) run with
>>>>>> open boundaries. The domain is the Drake Passage. A plot of
>>>>>> bathymetry and velocities can be found in
>>>>>> http://mitgcm.org/~mlosch/noise.png
>>>>>>
>>>>>> Whatever friction parameters I have tried (viscAh=1e0 to 2e1
>>>>>> and viscA4=1e8 to 4e10, and similar for diffusivities, I also
>>>>>> tried the Leith/Smagorinski variants), I seem to get noise in
>>>>>> the north western part of the domain. What worries me is, that
>>>>>> 1. The noise seems to propagate (compare day 149 to day 214
>>>>>> in the bottom panels of the figure)
>>>>>> 2. The noise seems to be mainly in the x-direction
>>>>>>
>>>>>> I use mom_fluxform. For mom_vecinv the problem is there, too.
>>>>>> I use USE_ISOTROPIC_SCALING (for viscosities because my y-
>>>>>> grid varies with y) and do not use COSINEMETH_III (although
>>>>>> that probably doesn't make much of a difference). When I turn
>>>>>> off USE_ISOTROPIC_SCALING, the noise is still there, but the
>>>>>> x- alignment is slightly less obvious (although very much
>>>>>> present). If the noise were deltaX in both directions, I
>>>>>> would be concerned about my friction parameters. Here I
>>>>>> suspect a problem in the viscosity implementation, but I
>>>>>> cannot see how and were.
>>>>>> There is some noise that is produced by the open boundaries,
>>>>>> but that usually goes away. I think that the generation of
>>>>>> the noise is connected to the topography around the tip of
>>>>>> South America
>>>>>>
>>>>>> I wonder if anyone has seen something like this before. What
>>>>>> do you think?
>>>>>>
>>>>>> Martin
>>>>>>
>>>>>> _______________________________________________
>>>>>> MITgcm-support mailing list
>>>>>> MITgcm-support at mitgcm.org
>>>>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> MITgcm-support mailing list
>>>>> MITgcm-support at mitgcm.org
>>>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>>>
>>>>
>>>> _______________________________________________
>>>> MITgcm-support mailing list
>>>> MITgcm-support at mitgcm.org
>>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>>
>>>
>>> _______________________________________________
>>> MITgcm-support mailing list
>>> MITgcm-support at mitgcm.org
>>> http://mitgcm.org/mailman/listinfo/mitgcm-support
>>>
>> --------------------------------------------------------
>> 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/ USA
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel
More information about the MITgcm-devel
mailing list