Newsgroups: comp.parallel.pvm
From: donato@cs.utk.edu (June Donato)
Subject: Re: pvm_reduce
Organization: CS Department, University of Tennessee, Knoxville
Date: 23 Nov 1994 14:30:03 -0500
Message-ID: <3b057rINNhc@thud.cs.utk.edu>

In article <3a82c2$2qkk@info2.rus.uni-stuttgart.de> zra60100@nfhsg2.rus.uni-stuttgart.de (Michael Resch) writes:
...
>a question concerning pvm_reduce: 
>how does it collect the data??? i mean, how does it send messages???
...
>michael

The initial implementation of pvm_reduce is very simplistic.
Each node sends its data to the specified host node of the reduce 
operation.  As the host receives the data from each node it performs 
the (binary type - add, product, min, max, ...) operation
on the newly received data and the already updated (or original)
data it has.  Certainly, this could be optimized, but at some
expense to robustness.  If one node goes down, the host
will still receive the data from all the other nodes, and thus
as much partial information as possible is still made available.
However, if a tree type algorithm were used and a node were to 
leave the group (or go down), then data from any of its children 
would be lost also.

Future releases of pvm will make use of native operations on
between nodes on parallel architectures (such as the IPSC/860,
Paragon, CM-5, SP-2, ...).

 /-------------------------------------------------------------------------\
|*   Dr. June M. Donato           phone: (615) 576-7928                    *|
|    Oak Ridge National Lab       email: donato@msr.epm.ornl.gov            |
|    P.O. Box 2008, Bldg. 6012    FAX:    (615) 574-0680                    |
|    Oak Ridge, TN 37831-6367     WEB:    http://www.epm.ornl.gov/~donato/  |
|                                                                           |
|*   physical office:             ORNL, X-10, Bldg. 6010, room 108         *|
 \-------------------------------------------------------------------------/

