Hi,
Mike Zingale has told me about how the eint_switch can be dangerous
at it's default seetting, we saw this in a case of a sedov explosion
with very odd behaviour due to forced refinement. I have found that on
the Evrard collapse problem (a standard test for SPH codes) this same
problem arises. The Evrard problem is a spherical distribution with 1/r
density profile and a cold temperature which collapses with gravity. I
have attached to this mail a tarball of the setup directory I used. The
problem is described in 1988MNRAS.235..911E and 1993A&A...268..391S
(those are adsabs bibcodes).
I ran this problem with eint_switch=1e-4 and have plotted the energy, as
outputted in flash.dat here:
http://imp.mcmaster.ca/~colinm/FLASH/evrardenergy.png
Note the massive energy loss, and resulting crash of the code.
I also ran this problem, with eint_switch=0 and plotted the same, here:
http://imp.mcmaster.ca/~colinm/FLASH/evrardenergy2nde0.png
That's about a 7% gain.
The difference between these two suggest again that something is wrong
with the dual energy formulation used in FLASH. My collegue (James
Wadsley) noted that it's worrying that a divergance is calculated
seperatly in each direction of the PPM update in update_soln.F90.
Also, the results for this test are rather poor in their spherical
symmetry. I have a scatter plot of the cell densities by radius here:
http://imp.mcmaster.ca/~colinm/FLASH/evrarddensscatter2nd08.png
Ideally, of course, this would be a thin curve instead of being so
smeared. Do you have any recommendations that might improve the
performance of this test, while not doing anything unrealistic for a
cosmology-type simulation (i.e. can't use spherical coordinates). SPH
codes can give much better resultsd than this for this problem, is there
any particular reason why this test might be "unfair" towards AMR
techniques?
-Colin McNally
McMaster University
This archive was generated by hypermail 2b30 : Wed Aug 28 2002 - 16:13:23 CDT
