[mpi3-coll] Our MPI-2.2 Items

Torsten Hoefler htor at cs.indiana.edu
Sun Nov 9 11:14:09 CST 2008

Hello workgroup,
our WG has several open MPI-2.2 items. I am listing them here for
discussion. We should make progress on all of them (i.e., bring them to
e review-ready state or request more feedback from the forum). 

We should discuss the remaining open items in the mailinglist and/or
during our next teleconference (the Thursday after SC08). Here are the

#23 (Generalized v-collectives: allow unknown recvcounts / rdispls):
 - seems to have several problems and the benefits seem to be very limited
  (main concerns: benefits only exist for small messages and an upper
  bound to the cumulative message size has to be known before the
  operation. So the user could as well do a non-vector operations with
  slightly larger buffers and achieve higher performance with negligible
  memory overhead. It is also unclear how to implement this performing
  (e.g., with a tree algorithm) without excessive memory copies and
  forcing serial receives.)

#24 (Add a local Reduction Function):
 - seems to be ready for first reading
 - Jesper volunteered to add required changes to document, I will review

#27 (Regular (non-vector) version of MPI_Reduce_scatter):
 - is missing a use-case, implementation and a clear benefit

#30 (Clarification to intercomm MPI_Barrier):
 - awaiting reviews

#31 (Add MPI_IN_PLACE option to Alltoall):
 - it's still unclear whether we want MPI_IN_PLACE or MPI_Alltoall_replace()
 - we should have a straw-vote during the next telecon (after SC) and
   update the proposal

#33 (Fix Scalability Issues in Graph Topology Interface):
 - issue if we want weights (and what weights) or not is open
 - otherwise it seems ready for a first reading
 - we should come to a conclusion with the weight issues before proceeding
   and update the proposal

#38 (Matched probe/receive):
 - seems to fit better in p2p WG
 - will move over

#44 (Non blocking versus non-blocking versus nonblocking):
 - trivial, ready for reviews by chapter authors

#26 (Add a callback function if a request completes):
 - relatively clear "no" for 2.2
 - will move to p2p WG

New additions are still welcome!

All the best,

 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler       | Postdoctoral Researcher
Open Systems Lab      | Indiana University    
150 S. Woodlawn Ave.  | Bloomington, IN, 474045, USA
Lindley Hall Room 135 | +01 (812) 855-3608

More information about the mpiwg-coll mailing list