Transport protocols and bulk file transfer

[Scott Brim <>]

I'm part way into writing a small essay on this very subject, figuring
out what research to foster :-).  I probably agree with Larry on all
points.

The best thing about their scenario and multicast is that they have a
lot of fan out - 30 sites well distributed - and they have predictable
consumption.  It's not like that for most research datasets, where a few
sites, far apart, want a particular dataset only once or twice.
Reliable multicast could apply.  Their problem would be finding
equipment that supports it.

I can think of five types of mechanisms to handle staged one-to-many
distribution in general: (1) store data in case anybody wants it and
forward on demand; (2) store only when it's been asked for and forward
immediately; (3) intermediate layer dynamic streaming like Phoebus; (4)
L3 multicast; and (5) "lower layer" (SDN, whatever) multicast.

In brief I think 1 isn't very applicable (we're not Netflix) and 4 has
reliability/retransmission issues as Larry pointed out (it's a great
research topic).  2 is akin to hierarchical distribution and can be very
useful.  3 has already demonstrated its worth in Phoebus.  As for 5 (SDN
trees), well, you would have the same reliability issues, but you could
handle them with unicast side channels.  #5 essentially takes the
application-layer hierarchical distribution and moves it down into the
SDN infrastructure - I think it's worth investigating.

Scott