[Beowulf] Alternative to MPI ABI

Patrick Geoffray patrick at myri.com
Fri Mar 25 11:47:44 PST 2005

Hi Don,

It looks a lot more like a runtime environement (care about resources, 
scheduling, failover) than a programming model to me. MPI is much more a 
programming model than a runtime environement, so I don't think they are 
that orthogonal.

Donald Becker wrote:
>    MPI has a model of initialize-compute-terminate.
>      There is no explicit support for checkpointing, executing as a
>      service, or running "forever".

There is no explicit support but people have been checkpointing parallel 
jobs for a while. You just need to be flush all pending communications: 
pass a token a couple of times and you will have a clean cut.

>    MPI's strength is collective mathematically-oriented operations, not
>      communication.  I understand that even the name "Message Passing.."
>      indicates that stream communication isn't the focus, but many
>      applications expect and work well with a sockets-based model.

Aaargh. Sockets are definitively not a programming model suited for 
parallel codes. Client/server, maybe, not tighlty coupled applications. 
Look at the bodies left from trying to do zero-copy and OS-bypass 
sockets: when you apply all of the constraints, you basically gut out 
the stream paradigm.

>    Communicators besides MPI_COMM_WORLD are rarely used.  The capability
>      adds complexity with little benefit.

Most libraries use communicators for isolation. Look at the BLACS 
contexts for example.


Patrick Geoffray
Myricom, Inc.

More information about the Beowulf mailing list