[mpi3-coll] NBC Draft Revision 5

Adam Moody moody20 at llnl.gov
Wed Feb 25 13:29:17 CST 2009

Hello all,
I opened a ticket to add the text to the collective intro regarding 
access restrictions on MPI_IN_PLACE like we have in the NBC proposal:
Erez reviewed it and pointed out that it's not really necessary.  All 
collectives specify MPI_IN_PLACE should be applied to the receive buffer 
which has the stronger constraint, except for the root in scatter and 
scatterv which does not receive anything.  Then, after I looked over 
things again, I realized that my current statements are actually more 
restrictive than they need to be for scatter{v}.

    "When using the "in place" option, message buffers function as both 
send and receive buffers. Such buffers should not be modified or 
accessed until the operation completes."

Since nothing is received at the root in scatter{v}, it could in fact 
issue several NBC scatter{v} calls on different communicators (reusing 
an active send buffer).  To fix this, we have two options;
    1)  Change "When using the 'in place' option, message buffers 
function"  to  "When using the 'in place' option, receive message 
buffers may function"
    2)  Close ticket 131 and strike the similar lines in the NBC 
proposal -- just rely on the implicit send/receive buffer constraints
I'm in favor of option #2 myself, since I'm not sure how much info #1 adds.

If we go with option #2, it would leave a single-sentence paragraph on 
page 51, line 11 in the NBCv5 text.  We could append this sentence to 
the paragraph above it.

Any opinions?

Torsten Hoefler wrote:

>Hello workgroup,
>I just posted revision 5 of the NBC draft to ticket #109.
>The draft is also available at [1] and a diff to revision 4 at [2].
>[1]: http:// www. unixer.de/sec/nbc-proposal-rev-5.pdf
>[2]: http:// www. unixer.de/sec/nbc-proposal-rev-5.diff
>Please review and comment!
>All the Best,
>  Torsten

More information about the mpiwg-coll mailing list