[Developers] optimising the size of checkpoints

[Christian David Ott]

On Tue, 14 Jun 2005, Thomas Radke wrote:

> Currently the checkpoint methods available in Cactus save all the timelevels 
> of all grid variables which have storage allocated, resulting in a 
> maximum-size checkpoint.
> At least for some variables (eg. NaNChecker::NaNmask or other analysis grid 
> functions), checkpointing doesn't seem necessary; other variables, especially 
> scalars (eg. IOBasic::next_info_output_time), can be easily recomputed after 
> recovery, thus saving another item in the checkpoint plus its attached 
> metadata (which in this case would be larger than the actual data).
>
> I propose to tag such grid variables with something like 'CHECKPOINT = no' in 
> a thorn's interface.ccl. The checkpoint routines could then query the tags 
> table and save only non-tagged variables.
> Of course, one would also have to add the logic to properly initialise tagged 
> variables in CCTK_POST_RECOVER_VARIABLES if need to.
>
> Comments ?

Hi,

I think this is a great idea. I just looked (via h5ls) through one of my 
checkpoint files and found that ~30-40% of the variables in there just 
don't need to be checkpointed. MoL error estimates for example could be 
reset after at the time of recovery.


  - Christian