Newsgroups: comp.parallel.mpi
From: James Cownie <jcownie@dolphinics.com>
Subject: Re: [Q] MPI_Bcast on Ethernet
Organization: Dolphin Toolworks
Date: Thu, 08 May 1997 09:19:59 +0100
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <33718CAF.27C1@dolphinics.com>

Johann Leichtl wrote:
> 
> hi,
> 
> if i'm not mistaken, a mpi (the same is true for pvm) broadcast consists
> of a number of point to point messages, only that mpi takes care of it.
> you get the same effect as if you run a loop and send everybody in the
> group the message.
> 
> obviously this has some performance drawbacks.

In MPI we took a lot of trouble to ensure that broadcast is correctly
implementable using a tree, or (if the underlying medium supports it)
some other truly broadcast implementation. If your MPI implementation
does broadcast in time linear in the size of the communicator you should
complain to the implementor.

Enabling such optimisations is the reason that MPI has a broadcast,
rather than a multi-send.

-- Jim 

James Cownie 
Dolphin Interconnect Solutions
Phone : +44 117 9071438
E-Mail: jcownie@dolphinics.com