[Beowulf] Broadcast - not for HPC - or is it?
matthurd at acm.org
Fri Sep 24 03:21:55 PDT 2010
I'm associated with a somewhat stealthy start-up. Only teaser product
with some details out so far is a type of packet replicator.
Designed 24 port ones, but settled on 16 and 48 port 1RU designs as
this seemed to reflect the users needs better.
This was not designed for HPC but for low-latency trading as it beats
a switch in terms of speed. Primarily focused on low-latency
distribution of market data to multiple users as the port to port
latency is in the range of 5-7 nanoseconds as it is pretty passive
device with optical foo at the core. No rocket science here, just
convenient opto-electrical foo.
One user has suggested using them for their cluster but, as they are
secretive about what they do, I don't understand their use case. They
suggested interest in bigger port counts and mentioned >1000 ports.
Hmmm, we could build such a thing at about 8-9 ns latency but I don't
quite get the point just being used to embarrassingly parallel stuff
myself. Would have thought this opticast thing doesn't replace an
existing switch framework and would just be an additional cost rather
than helping too much. If it has a use, may we should build one with
a lot of ports though 1024 ports seems a bit too big.
Any ideas on the list about use of low latency broadcast for specific
applications in HPC? Are there codes that would benefit?
More information about the Beowulf