🚀 go-pugleaf

In article <1992Aug11.121234.28784@hubcap.clemson.edu> jek@cs.duke.edu (James E. Kittock) writes:
>
>This summer, I have been working with PVM on a project to parallelize
>codes that simulate convection.  The parallelization is being done
>through domain decomposition, i.e. each of N subprocesses works on
>1/Nth of the original domain.
I am curious: when you partition, are the partitions of the array 
in contiguous chunks, or if they are interleaved, are the chunks
reasonably large? (say a multiple of page size). 

If each process's 1/Nth of the array is in a contiguous chunk,
then it seems to me you could use a DSM. 
When a process is scheduled, it uses its process number 
to compute a pointer to its chunk of work and does the computation. 
That is basically what I have done here on the ELC farm for 
several applications. Then you are using PVM for scheduling and
the DSM for sharing data.

Even supposing they aren't contiguous, or they aren't aligned
on nice boundaries, if you are using a DSM then each 
process can do assignment statements from the global array to the small
local copy. The programming effort is easy, at least for the applications
I have tried. 

the put_automatically thing you propose looks difficult. 
I don't want to write it either :-)

ron
-- 
--------------------------------------------------------------------------------
satisfaction: the feeling you get when you        | rminnich@super.org
run newfs on a DOS disk so you can install 386BSD.| (301)-805-7451
Now if only I could do this to AIX on the RS/6k ...|