Newsgroups: comp.parallel.mpi
From: colella@unwanted.nimue.plk.af.mil
Subject: MPI Probes
Organization: Air Force Phillips Lab.
Date: 10 Aug 1998 19:25:38 GMT
Message-ID: <6qnhfi$q5c$1@pr1.plk.af.mil>

Hi,

If I'm not sure how large a message is that needs to be received, and I do a blocking
or non-blocking probe, where is the message if the probe returns true?  Is it still
on the sending node?  Or is it buffered somewhere on the receiving node?  Is this
platform dependent?  I don't see anywhere in the standard where this is discussed,
so I suppose it must platform/implementation dependent.  I'm looking specifically
at the IBM MPI implementation for the SP2.  We're getting very erratic behavior,
and all of the receives for messages large and small, are preceeded by either
MPI_IProbe or MPI_Probe, and a blocking receive is posted if and only if the probe
returns true.  I'm wondering if the receiving node may be running out of space, or 
the sending node, since some of the messages can be quite large (in the MBytes), and
the nodes may be pretty far out of synch.

I'd like to do these sends and receives in a more "static" way, getting rid of the 
probes.  I believe we'll see more consistent behavior if I do this, but I'd like to 
know if I'm barking up the wrong tree before I put out this kind of effort (and it 
will be an effort!).  If this is discussed in the standard anywhere, and I missed it,
please tell me.

Thanks!

shari
colella@unwanted.nimue.plk.af.mil

remove the unwanted for email