[MPICH] Can I ask MPI to not copy the data to shared-mem ?

[chong tan]

Agree.  As of this moment, I am looking for some way to squeeze the last drop of perofrmance for my application.  I do not use Isend or Irecv, and that actually slow down my application in some situation.
   
  tan
  

Jean-Marc Saffroy <saffroy at gmail.com> wrote:
  On Tue, 9 Oct 2007, chong tan wrote:

> having both sender and reciever being the same shared memory may 
> require Barriers to be placed to block writing to the mem before it is 
> completed consumed. That may actually slow done the run.

MPI_Isend could still benefit from this approach, eg:
- process A does Isend (or Irecv), which simply marks the message to be 
sent (or received), then continues with useful computation
- then process B does Irecv (or Isend), which detects the mark and does 
the actual (and single) message copy, then marks the message as sent
- then processes A and B eventually do a Wait or Test

It will probably depend on how much the application allows communications 
to overlap with computation (but if it uses Isend/Irecv then it probably 
cares already?).


-- 
saffroy at gmail.com


       
---------------------------------
Need a vacation? Get great deals to amazing places on Yahoo! Travel. 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20071009/372a0ca7/attachment.htm>