Re: tiling issues

Rajeev Thakur (thakur@mcs.anl.gov)
Wed, 16 Apr 1997 11:38:07 -0500

> Date: Wed, 16 Apr 1997 09:20:27 -0700 (PDT)
> From: Linda Stanberry <linda@k2.llnl.gov>
>
> It has never been my interpretation that the tiling by a datatype was
> different for filetypes than for buftypes. Our implementation does
> not make any such distinction. Can you point to wording in Chapter 10
> that lead you to this conclusion?
>
> Linda

The definition of view on pg 205 and the illustration in figure 10.1
says that the filetype defines a pattern that is repeated to form the
view.

To me, the derived dataype (int,4) defines the pattern
(hole, int). Therefore, the file view will be
hole, int, hole, int, ...

I think this is fine; that was the intent all along.

But if the same datatype is used in MPI_Read, it means the buffer will
contain count "contiguous" copies of type datatype. Contiguous is
defined in terms of extent of the datatype (pg. 73, MPI-1), which for
the above is four. So, the memory layout will be

hole, int, int, int, ...

I think we need to define file tiling more precisely. The displacement
between consecutive filetypes is not the "extent" as in memory, but
"extent + lb(filetype)".

Rajeev