Newsgroups: comp.parallel.pvm
From: rztpj@mufus.rz.tu-harburg.de (Peter Junglas)
Subject: Re: problems with pvm_reduce
Organization: Techn. Uni. Hamburg-Harburg, Germany
Date: 15 Jul 1994 13:21:11 GMT
Message-ID: <3062g7$e1s@rztsun.tu-harburg.de>

I had the same problems with pvm_reduce, until I added a pvm_barrier call after the pvm_joingroup and before the pvm_lvgroup. Now everything works fine and the reason for the strange behaviour seems clear:

Since the reduce doesn't include a barrier, it looks up the group members at the time of the call. But then some of them may have already left the group, after having done a reduce only within their active subgroup.  It even happened that the whole sequence "joingroup, reduce, lvgroup" was executed sequentially on one host after the other, so that everyone computed a reduce just with his own single value. Often the reduce produced an error, because the group changed during the reduce itself. 

I think, lots of barriers are the price one has to pay for the dynamical group concept of pvm. Probably this is the reason, why many people dislike the idea of dynamical groups and why MPI only provides static groups. I'd really like to know the opinion of the net about this topic: Do we need dynamical groups at all?

  Peter

