Welcome, Guest
Username: Password: Remember me
HELPDESK

Here we can describe more what should be posted here

TOPIC: FullPos high to low resolution

FullPos high to low resolution 2 months 3 weeks ago #2568

Dear all,

The incremental approach of Harmonie_Arome 4dvar requires low resolution versions of the first guess. This is accomplished (see scr/4DVprolog) by running E927 with as input file the high resolution first guess and by specifying new NLON's etc by

NLON=$(( $nlon / $res ))
NLAT=$(( $nlat / $res ))
EZONE=$(( ($ezone + 1)/$res - 1 ))
NDLUXG=$(( $NLON - $EZONE ))
NDGUXG=$(( $NLAT - $EZONE ))

For the default settings with a resolution of 2500m and res=3 or 6 this has never led to issues.

But now using a resolution of 500m I am unable to get a low resolution (I tried res=3,5,6,and 10)
Often with the message that the outfile produced by E927 is outside the infile domain. Also reducing NLON and NLAT substantially does not help.

If anyone has a clue, that would be great!

Thanks,
Jan

FullPos high to low resolution 2 months 3 weeks ago #2569

  • Ole Vignes
  • Ole Vignes's Avatar
  • OFFLINE
  • Administrator
  • Posts: 43
  • Thank you received: 12
Hi,

for this to work your NLON and NLAT must of course be divisible by
all your ILRES ($res) factors. But in addition, the remainder must also be divisible by 2.

What are your NLON and NLAT?

Also, I remember from 10 years back or so that we had to change a domain test to accept an "epsilon" difference in domain borders because of accuracy issues in computing the new domain at new resolution(s). Perhaps the epsilon needs to be reconsidered for your case?
Also, if you can point me to some logs (e.g. at cca) I could have a closer look.

Ole
The following user(s) said Thank You: Jan Barkmeijer

FullPos high to low resolution 2 months 3 weeks ago #2570

Yes Ole I am aware of the 235 rule and checked the domain (also after the resolution change) in addition with the wiki domain tool (for which one has to log in again ...)

The exp is nl/nlf/4DVAR_500_NOROT750

It may be that this epsilon plays up with the 500m resolution.

Thanks,
Jan

FullPos high to low resolution 2 months 3 weeks ago #2571

  • Ole Vignes
  • Ole Vignes's Avatar
  • OFFLINE
  • Administrator
  • Posts: 43
  • Thank you received: 12
With NLON=NLAT=750=2*3*5*5*5 I don't think factors of 6 or 10 can work, because if you divide by any of those you have no factor 2 left.
Why 3 and 5 don't work I'm not sure about.
Your crash is in aladin/fullpos/suefpg3.F90 line 250 it seems, which is an abort. What I would do is to put just before the call to abor1 this line:

CALL FLUSH(NULOUT)

since now I think we lose some essential information about how far off the offending points (outside area) are.

FullPos high to low resolution 2 months 3 weeks ago #2572

Just gave a flavour of the various settings I tried, including rotated and nonrotated domains.

I will now follow your suggestion and put a flush

FullPos high to low resolution 2 months 3 weeks ago #2573

The flush results in the same output that was already in the NODE file (see below). For some reason it seems that the OUT domain is shifted by 500m, the resolution of the IN domain. The info is for original NLON=NLAT=750 and resolution factor=5 (so OUT domain has resolution 2500m). With factor to 2 the negative values become order E-10. I started to look at ways to manipulate the X and Y, but I am sure there are more elegant approaches.



SUFPG2 FOUND OUT : LFPOSHOR = T
X MIN = 0.0000000000000000 X MAX = 369500.00000000000
Y MIN = 0.0000000000000000 Y MAX = 369500.00000000000
INDEX, X, Y = 1 -500.00000000002910 -500.00000000000000
INDEX, X, Y = 2 1999.9999999999709 -500.00000000000000
INDEX, X, Y = 3 4500.0000000000000 -500.00000000093132
INDEX, X, Y = 4 6999.9999999999709 -500.00000000000000
INDEX, X, Y = 5 9499.9999999999709 -500.00000000000000
INDEX, X, Y = 6 12000.000000000029 -500.00000000093132
INDEX, X, Y = 7 14499.999999999971 -500.00000000000000
INDEX, X, Y = 8 17000.000000000000 -500.00000000093132
INDEX, X, Y = 9 19500.000000000000 -500.00000000000000
INDEX, X, Y = 10 21999.999999999942 -500.00000000000000
ETC ...
SUEFPG3 : OUTPUT POINTS HAVE BEEN DETECTED OUTSIDE THE INPUT MODEL DOMAIN
ABOR1 CALLED
SUEFPG3 : ABOR1 CALLED

FullPos high to low resolution 2 months 2 weeks ago #2578

As an alternative to fullpos I also tried the option 4DVprolog_gl_only. After replacing $BINDIR/gl_grib_api by $BINDIR/gl the execution of the script leads to a core without an obvoius error message. Perhaps this option has never been tested before? Any clue would be helpful.

FullPos high to low resolution 2 months 2 weeks ago #2579

  • Ulf Andrae
  • Ulf Andrae's Avatar
  • OFFLINE
  • Administrator
  • Posts: 308
  • Thank you received: 35
Jan,

It has been tested but didn't work very good ( for reasons unknown to me) so it's not maintained. I suggest that it's removed as an option.

Ulf

FullPos high to low resolution 2 months 2 weeks ago #2580

Thanks Ulf, and given your expertise on gl, I have little hope to crack this gl alternative (albeit it is merely a 're-numbering' problem as there is no vertical interpolation involved). I conclude that I am back in the fullpos world.
Time to create page: 0.066 seconds