H5Part development and discussion

[Mark Howison <MHowison AT lbl.gov>]

Hi Achim, you are right that it isn't a huge amount of memory (there
are actually 2 * 6 * nprocs because it stores both a "user" and
"dissolved" layout), so I will just leave it as is for now. Mark

On Fri, Jul 9, 2010 at 7:01 AM, Achim Gsell <achim.gsell AT psi.ch> wrote:
> Hi Mark
>
> On Jul 8, 2010, at 5:54 PM, Mark Howison wrote:
>
>> I'm still working through the H5Block code, and I wanted to find out
>> more about the original design decisions for handling arbitrary ghost
>> zones, which introduces alot of complexity in "dissolving" them. I
>> don't see an easy way to get around the O(n) space required in the
>> H5Block file struct to store every proc's field layout for use in the
>> ghost zone routines.
>
> I implemented arbitrary ghost zone on request of LBNL ... But is the O(n) 
> space required really an issue? The layout occupies 
> 6*sizeof(h5_int64_t)*nprocs bytes per process. With 64k cores it's just 3MB 
> (or 0.0022% of the available memory we have on Franklin/core).
>
>> Do we need to support arbitrary ghost zones?
>
> Good question, actually I don't know. As far as I know, we don't need them 
> at PSI.
>>
>
> Achim
>
>