Re: [h5md-user] Update on H5MD from Stuttgart

[Peter Colberg]

On Sun, Jul 21, 2013 at 10:43:03PM +0200, Pierre de Buyl wrote:
> I pushed modifications and found a surprise issue. 
> 
> In the box specification, one stores edges and offset as attributes for time
> independent boxes. This goes against the new idea that any data is a dataset 
> or
> a H5MD group. Still, datasets make sense in "/trajectory" (for species or
> masses, say) but attributes are more appropriate for a box edges and offset.

I think this can be resolved without any further changes.

While I support the new idea to use datasets in place of step/time
groups for the case where a dataset is time-invariant, it should not
be applied in a dogmatic sense, e.g., against good implementation
practice.

A good practice is to store data significantly smaller than 64 KiB in
attributes, and larger data in datasets. (HDF5 library version 1.8 or
later does support attributes larger than 64 KiB [1], but this
requires special handling.)

[1] http://www.hdfgroup.org/HDF5/doc/H5.format.html#AttributeMessage

I suggest to limit the use of attributes to the case where data will
always be significantly smaller than 64 KiB. This includes the case
of time-invariant box edges and offset data.

For cases where data could potentially reach or exceed 64 KiB,
datasets should be used. This includes the case of time-invariant
trajectory data, which scales with the system size.

Peter