[Carpet] [Carpet-darcs] CarpetIOHDF5: more efficient recovery with same number of processors

[Erik Schnetter]

On Feb 13, 2006, at 11:37:37, Thomas Radke wrote:

> Erik Schnetter wrote:
>> Thomas,
>>
>> please excuse the stupid question.  What happens if the number of
>> processors is the same, but the grid hierarchy is distributed in a
>> different manner across the processors?  This can happen if e.g. the
>> grid hierarchy is distributed manually, of if the recovering process
>> uses a different grid hierarchy.  There should be a check that the
>> current processor's grid hierarchy can be found completely in this
>> single file.
>
> Hi Erik,
>
> your question isn't stupid at all. Now that I think of it, my  
> assumption
> for this optimisation patch isn't strong enough. As you say, if the  
> grid
> topology changes while keeping the same number of processors the  
> current
> recovery code will not work properly.
>
> Is it enough to save the processor topology in the checkpoint files
> (that's what CactusPUGHIO/IO{HDF5,FlexIO} do) ? Or do I also need to
> save the entire grid setup ?

I think that each processor can make the decision whether to read one  
or more file independently of the others.  It could start to read  
everything from the file, and if there is part of its grid hierarchy  
uninitialised afterwards, it has to look into the other files.   
(Isn't there already a flag or a bbox set that says how much is not  
yet initialised?)  Once this set is empty, it can stop opening other  
files.

If that works, this criterion is independent of the number of  
processors and the number of files.

Isn't that check already there?  This is the if (all_read_completely)  
check in line 384 of Input.cc.

Are we still doing stupid things with regard to synchronisation?  Are  
we reading the ghost zones in as well?  Maybe we shouldn't, and we  
could synchronise afterwards instead?  Or do we want to support  
unsynchronised grid functions?  Maybe we want to synchronise inter- 
processor boundaries, but not prologation boundaries?  I remember  
twisted discussions.

-erik

-- 
Erik Schnetter <schnetter at cct.lsu.edu>

My email is as private as my paper mail.  I therefore support encrypting
and signing email messages.  Get my PGP key from www.keyserver.net.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 186 bytes
Desc: This is a digitally signed message part
Url : /archives/developers/attachments/20060213/b3b4dc27/attachment.pgp