[Developers] Recover SaveAndRestore variables with PUGH

Luca Baiotti baiotti at aei.mpg.de
Wed Oct 18 03:02:02 CDT 2006


Hi,

there is a problem recovering with PUGH a simulation in which some
variables ended up to be registered to MoL as SaveAndRestore. (This
happens for example using whisky and admbase::evolution_method = none)

I had email exchange with Ian Hawke about this and he replied:

> I believe that the past timelevels of the metric are not set
> correctly. This is because they are not checkpointed by PUGH. If you
> hacked MoL to run the InitialCopy routine at PostRecover, I expect you
> would see the answer you expected.
>
> I am not sure how to get around this, other than by asking PUGH to
> checkpoint more timelevels...


I tried what Ian first suggested, that is I modified the MoL
schedule.ccl as follows:


if (initial_data_is_crap)
{
   if (copy_ID_after_MoL_PostStep)
   {

     schedule MoL_FillAllLevels AT CCTK_POST_RECOVER_VARIABLES AFTER
MoL_PostStep
     {
       LANG:C
     } "A bad routine. Fills all previous timelevels with data copied
from the current."


     schedule MoL_FillAllLevels AT PostInitial AFTER MoL_PostStep
     {
       LANG:C
     } "A bad routine. Fills all previous timelevels with data copied
from the current."
   }
   else
   {

     schedule MoL_FillAllLevels AT CCTK_POST_RECOVER_VARIABLES BEFORE
MoL_PostStep
     {
       LANG:C
     } "A bad routine. Fills all previous timelevels with data copied
from the current."

     schedule MoL_FillAllLevels AT PostInitial BEFORE MoL_PostStep
     {
       LANG:C
     } "A bad routine. Fills all previous timelevels with data copied
from the current."
   }
}


and indeed it worked.

Should we introduce a parameter to have this InitialCopy scheduled at
post_recover_variables for PUGH?

And Ian replied:

> I think some variant of this is needed, yes. (In fact the correct
> thing to do for MoL should be only to copy the SaveAndRestore
> variables - there should not be a problem for evolved variables with
> PUGH. Of coures, running the whole routine shouldn't give problems, I
> think...).



More ideas?

Ciao,

Luca.



More information about the Developers mailing list