Date: Wed, 04 Jan 2006 08:13:22 -0600
To: Lars Schneidenbach <lschneid@XXXXXXXXXXXXXXXXX>
From: William Gropp <gropp@XXXXXXXXXXX>
Subject: Re: One-sided communications example issue
Cc: mpi-comments@XXXXXXXXXXXXX
In-Reply-To: <200601041421.36521.lschneid@XXXXXXXXXXXXXXXXX>
References: <200601041421.36521.lschneid@XXXXXXXXXXXXXXXXX>
Mime-Version: 1.0
Content-Type: text/plain; charset="iso-8859-1"; format=flowed
Sender: owner-mpi-comments@XXXXXXXXXXXXX
Precedence: bulk

At 07:21 AM 1/4/2006, Lars Schneidenbach wrote:
text under discussion: Example 6.10, page 135/136

Hi,

while digging into details of MPI2 one-sided communications, I tried to
understand the examples (Section 6.5).  I think there's an issue with
example 6.10 (page 135):

At page 136 Line 24 the code runs a check on A1 inside the access epoch
of win1. As far as I understand, the content of A1 is not correct until
MPI_Win_complete(win1) returns in line 26.

I had a look to the errata-documents and found nothing about that.

I think the example does not work the way it should work. Can you tell
me if I'm right or wrong?

I believe that you are correct, though if the definition of "converged" includes only the parts updated by Update and not Update2, the code will work as written, though that is probably not what is intended.  In real code, you wouldn't call "converged" twice like this (once in the if and once in the while); you are more likely to use a flag variable to indicate whether to call post.

Bill


Regards
Lars Schneidenbach

--
 Dipl.Inf. Lars Schneidenbach         lschneid@XXXXXXXXXXXXXXXXX
 Universität Potsdam                      Tel : +49 331/977 3123
 Inst. f. Informatik, Postfach 90 03 27   Fax : +49 331/977 3122
 D-14439 Potsdam, Germany


William Gropp
http://www.mcs.anl.gov/~gropp