availability of Memory compression routine

William Gropp gropp at mcs.anl.gov
Fri Jul 19 06:58:01 PDT 2002

At 12:13 AM 7/18/2002 -0500, Dean Johnson wrote:
>As for the communications piece, isn't an mpi_broadcast just a loop over
>a send to the available nodes (ie. a mere convenience).

No implementation of MPI_Bcast that I know of is this simple.  All use at 
least a fanout tree (with log p fanout); more sophisticated versions use 
(for large data) scatter/gather operations that make full use of a switched 
network, and very sophisticated versions use a suite of algorithms, 
choosing the one to use based on the network and message size .  Multicast 
is tricky, since (as you note) basic multicast isn't reliable and not all 
networks support it, but there have also been some efforts to provide 
implementations of MPI_Bcast that use IP multicast as well.


More information about the Beowulf mailing list