[cgl_discussion] TIPC

Andy Pfiffer andyp at osdl.org
Thu Feb 13 14:16:58 PST 2003


On Thu, 2003-02-13 at 13:04, Stephen Hemminger wrote:
> How does TIPC compare to MPI?  http://www-unix.mcs.anl.gov/mpi/

Short answer: different solutions to somewhat different problems, but I
could imagine building an MPI implementation on top of what is described
in the document.

MPI needs to concern itself with marshalling and unmarshalling data,
coexisting gracefully on lowest-reasonable common demoninator platforms
(posix-like systems with sockets), minimizing latency, and maximizing
bandwidth.  Strictly speaking, MPI doesn't need to worry too much about
cuts in the network or failure cases of one or modes within an
application.

I suppose that a TIPC "zone" is somewhat similar to an MPI "rank", with
the follow-on observation that messages are delivered to a queue
identified by triples:

	TIPC: { zone, service [, processor] }
	MPI:  { appid, rank, slot* } 

rather than, say {netdevice identifier, port number} like TCP/UDP.

*I haven't looked under the covers of an MPI implementation in years,
but last I knew, some implementations really wanted to set up a full set
of N*M connections for all-to-all communication.

Andy





More information about the cgl_discussion mailing list