>This really gives some credit to the theory that the primary purpose of television is to pacify people and have them forget the real problems they face.
There is plenty of room for making good use of
multicasting in high performance computing.
However, most clusters really don't use it
because either A) The switches don't support
it (other than 1-to-all) or B) TCP/IP doesn't
support it.
The message passing libraries probably being used on your cluster (either PVM or some variant of an MPI implementation probably have multicasting statements, but they are in actuality probably
implemented as a sequence of normal one-one communications. Now if you wanted to implement
your own simulated multicast in the cluster, where
you used some kind of tree structure to forward
messages to all the right nodes, have at it,
but you'll probably have to modify your message
passing library...
Dan
If you can get a few Linux/BSD boxes, or at least dual boot a few, get a freely available message passing library (PVM, or an MPI implementation) and let the kids learn a little about message passing programs. Almost any algorithm gets a lot more challenging, and getting good performance gets harder still. You can add concepts like speedup, efficiency, and isoefficiency into the lectures. Kids can compete on performance. It's not that complicated, we've had some success working with a local high school and an 8 machine cluster here. Worst case, you can get a message passing library for Windows -- but I can't in good conscience recommend it;) Dan ------------------ Dan Stanzione Parallel Architecture Research Lab Clemson (Yes, the Beowulf Underground guys)
Slashdot Mirror
Venture Capitalist
>This really gives some credit to the theory that the primary purpose of television is to pacify people and have them forget the real problems they face.
Nah, we have religion for that...
There is plenty of room for making good use of
multicasting in high performance computing.
However, most clusters really don't use it
because either A) The switches don't support
it (other than 1-to-all) or B) TCP/IP doesn't
support it.
The message passing libraries probably being used on your cluster (either PVM or some variant of an MPI implementation probably have multicasting statements, but they are in actuality probably
implemented as a sequence of normal one-one communications. Now if you wanted to implement
your own simulated multicast in the cluster, where
you used some kind of tree structure to forward
messages to all the right nodes, have at it,
but you'll probably have to modify your message
passing library...
Dan
If you can get a few Linux/BSD boxes, or at least dual boot a few, get a freely available message passing library (PVM, or an MPI implementation) and let the kids learn a little about message passing programs. Almost any algorithm gets a lot more challenging, and getting good performance gets harder still. You can add concepts like speedup, efficiency, and isoefficiency into the lectures. Kids can compete on performance. It's not that complicated, we've had some success working with a local high school and an 8 machine cluster here. Worst case, you can get a message passing library for Windows -- but I can't in good conscience recommend it ;) Dan ------------------ Dan Stanzione Parallel Architecture Research Lab Clemson (Yes, the Beowulf Underground guys)