Skip to Content.
Sympa Menu

h5part - Re: [H5part] Re: H5Block and ghost zones

h5part AT lists.psi.ch

Subject: H5Part development and discussion

List archive

Re: [H5part] Re: H5Block and ghost zones


Chronological Thread 
  • From: Andreas Adelmann <andreas.adelmann AT psi.ch>
  • To: Mark Howison <MHowison AT lbl.gov>
  • Cc: Achim Gsell <Achim.Gsell AT psi.ch>, h5part AT lists.psi.ch, John Shalf <JShalf AT lbl.gov>
  • Subject: Re: [H5part] Re: H5Block and ghost zones
  • Date: Wed, 14 Jul 2010 01:20:27 +0200
  • List-archive: <https://lists.web.psi.ch/pipermail/h5part/>
  • List-id: H5Part development and discussion <h5part.lists.psi.ch>

Hi Mark ok now I understand.

In case of a stencil based computation with a very anisotropic computation
domain I can think of such cases. So we probable want to keep this feature
but maybe have a configure option to ride the more easy way; i.e. all the
dimensions have the same number of ghost nodes. This just in case of you
very well observed problem with O(1E6) cores.

What do you think ?

Best Amdreas



------
Dr. sc. math. Andreas (Andy) Adelmann
Staff Scientist
Paul Scherrer Institut WLGB/132 CH-5232 Villigen PSI
Phone Office: xx41 56 310 42 33 Fax: xx41 56 310 31 91
Phone Home: xx41 62 891 91 44
-------------------------------------------------------
Thursdays: ETH CAB H 83.1 +41 44 632 36 72
============================================
The more exotic, the more abstract the knowledge,
the more profound will be its consequences.
Leon Lederman
============================================



On 14.07.2010, at 00:46, Mark Howison wrote:

> Hi Andreas,
>
> It means you can have different depths of ghost cells on different
> sides of the 3D block. So the top might have +1 ghost cells, but the
> sides +2, etc. Adjacent blocks with ghost zones also overlap, and it
> is fairly complicated to "reduce" these into non-overlapping regions.
> Right now we handle the most general case, where we don't require any
> adjacency information from the user. There is an MPI_Allgather where
> each processor gets a copy of every other proc's field layout,
> requiring O(n) space. Then an O(n^2) algorithm runs redundantly on
> each proc to compute the reduction,
>
> It all works and I have never noticed any performance issues
> associated with it, but it may become a scalability problem in the
> future with say O(100K) or O(1M) cores.
>
> Mark
>
> On Sat, Jul 10, 2010 at 9:23 AM, Andreas Adelmann
> <andreas.adelmann AT psi.ch> wrote:
>>
>>>> Do we need to support arbitrary ghost zones?
>>
>>
>> Sorry I was maybe not following closely: what are rbitrary ghost zones
>> w.r.t. H5Block?
>>
>> AA





Archive powered by MHonArc 2.6.19.

Top of Page