[Beowulf] Broadcast - not for HPC - or is it?

Bogdan Costescu bcostescu at gmail.com
Tue Oct 5 11:38:40 EDT 2010

On Tue, Oct 5, 2010 at 3:40 PM, John Hearns <hearnsj at googlemail.com> wrote:
> However this device might be very interesting if you redefine your
> parallel processing paradigm.

I would say that's always true :-) If you can change the way you do
things, then adapting to the most appropriate hardware/software
combination makes sense to get the maximum of performance. Wait, this
has already happened... and then there was chaos... and then came MPI

> How about problems where you send out identical datasets to (say) a
> farm of GPUs.

You still need to send point-to-point data forth (what to do with the
bulk data) and back (the results). Then you have the big problem of
synchronizing the 2 channels: the broadcast one and the point-to-point
one. This can come from either GPUs taking different time to finish or
from point-to-point communication delays/congestion. And then the
efficiency brought by broadcast might just go away...

Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

This message has been scanned for viruses and
dangerous content by MailScanner, and is
believed to be clean.

More information about the Beowulf mailing list