Re: RFC: Coarray - implementation of the communication library

[Tobias Burnus <burnus at net-b dot de>]

On 06/01/2011 11:25 AM, Richard Guenther wrote:
> [Coarray implementation]
> Wouldn't it be easier to do a first implementation completely with
> blocking communication?  I will try to have a look at the non-blocking
> stuff as I have some (bad ;)) experience when teaching that to POOMA...

Well, at some point there has to be one-sided communication as that's 
the semantics of coarrays. Thus, either the called library does 
single-sided communication (e.g. via MPIv2, GASNet, ARMCI/Global arrays) 
- or one does it oneself using a combination of buffered and nonblocking 
communication. There might be some special cases where one can 
temporarily switch to blocking communication, but generally, it won't work.

Example:

if (this_image() == 3) then
  data[1] = data[2]
end if
sync all

Here, image 3 needs to tell image 2 to send data (to image 1 directly or 
to image 3) - and tell image 1 to receive data (from image 3 or directly 
from image 1). Thus, either one uses directly one-sided communication, 
or one sends (nonblocking) token around, telling each image what it 
should do. (The images need to wait anyhow for the "sync all" token from 
the other images.)
(For one reason against one-sided MPI v2, cf. [1], item 1.)

Tobias

[1] http://gcc.gnu.org/ml/fortran/2010-03/msg00201.html