Newsgroups: comp.parallel.mpi
From: Peter Sanders <sanders@mpi-sb.mpg.de>
Subject: Re: Bcast impmelmentation on SP2/mpich
Organization: University of Saarland, Computing Center, Germany.
Date: 23 Apr 1998 14:08:30 +0200
Message-ID: <ysyyawwlozl.fsf@mpi-sb.mpg.de>

gropp@godzilla.mcs.anl.gov (William Gropp) writes:

> In article <35350E75.C0C490D1@ecs.soton.ac.uk>,
> Panagiotis Melas  <pm95r@ecs.soton.ac.uk> wrote:
> >Does anybody know whether the broadcast implementation of MPI Bcast is
> >based on a binomial tree or not for the IBM SP2?
> 
> The MPICH implementation can be found in mpich/src/coll/intra_fns.c .
> The algorithm uses recursive subdivision; the root will send log_2(size of
> communicator) times.  The algorithm does not pipeline, nor is it tuned to 
> particular systems or message sizes.  Making the algorithms more 
> sophisticated has been on our list-of-things-to-do for some time, but 
> are currently lower priority to the MPI-2 implementation.
> 
> Bill

Our benchmark results at http://wwwipd.ira.uka.de/~skampi/
suggest that the IBM-SP switches to a pipelined 
algorithm for large messages.

Regards,

Peter Sanders