From: jonesbr@roger.ecn.purdue.edu (Brian R. Jones)
Newsgroups: comp.parallel.mpi
Subject: [Q] Out-of-order calls in MPI vs. PVM
Date: 25 Jun 1999 21:59:00 GMT
Organization: Purdue University, School of AAE
Message-Id: <7l0u34$q9g$1@mozo.cc.purdue.edu>
Xref: ukc comp.parallel.mpi:5250


Hello!

[Short version at the end]

Background:

I am in the process of converting a Fortran 77 3D CFD code from
PVM to MPI.  The ability to compile with PVM or MPI is a requirement,
thus I can not (extensively) modify the structure of the code.  I
realize that this limits the extent to which I can exploit the
added capability of MPI.  Eventually, PVM will be dropped and a
more effective MPI rewrite will occur.

That said, I currently have a direct translation completed (meaning
pvmfsend mapped to MPI_Send, etc.) and even managed to sidestep the
PVM spawning process in MPI.  I am having problems, however, with 
the data exchange between "blocks" of the 3D computational grid.
In the PVM version of the code, the process for each node looks
like this:

	1) perform computation on current block
	2) pass data (via pvmfsend/MPI_Send) at interface between
	   current block and neighboring blocks to corresponding
	   nodes.
	3) wait for interface data from neighboring nodes/blocks
	   (via pvmfrecv/MPI_Recv)

Note that there is no synchronization here.... if all the nodes had
an equal computational burden, they would all send their data at the
same time and then all receive data at the same time.

While I have had a bit of exposure to MPI, I'm new to PVM.  From what I
can tell, however, pvmfsend and pvmfrecv equate to MPI_Send and MPI_Recv
(i.e. both are blocking).  Dismissing that possibility, I'm now working
on why the program hangs.  I'm guessing that the PVM daemon is handling
the out-of-order communication that I will need to explicitly address
in MPI.  

The short version:  

A node has a message pending for receipt, but can't post the receive
until after sending off another message.  The buffering supplied by
the PVM daemon will allow this behavior, while the MPI code must be
structured to handle the MPI commands in the order that they occur.
Am I interpreting this correctly?  Is the simple (but ugly) solution
to add MPI_Probe commands before each send to test for a pending
receive?