In an event like that the normal tape selection goes like
        filling tape with node's data already on it
if not then
        scratch tape if "MAXSCR" for storage pool hasn't been hit AND
scratch tape still exists in the ATL
if not then
        any filling tape within the storage pool

when you add a new node, it goes through the same routine when its data goes
to a tapepool...

collocation in an environment with 2500 clients will cause a bunch of tape
mounts unless you limit the maxscr's
knowing that in a 3494 ATL (with 6+ frames) you are looking at an average of
90 seconds or so for each tape mount (until it starts writing)  and I'd
guess about the same for dismounts (from rewind to insert back into tape
storage slot)
so 2500 mounts/dismounts at 3 minutes each would put you at 2.08 hours for
just mounts & dismounts for migrating data from a diskpool to a tapepool and
if we are realistic, that many daily mounts would be hard on the loaders
within the tape drives...
Sure, without using collocation, eventually (from a statistic point of view)
each node would have some data on each tape and that would suck for a
restore...
as always it all depends...
I keep flip'n  environments back & forth trying to do the best for the
clients and unless you have lots of clients and lots of client data and only
a small part changes nightly and only if you end up doing LOTS of complete
restores, using collocation seems to be a flip of a coin (or at least in the
environments I deal with)
Now & again we end up doing a restore of an NT box with 30 GB and that takes
a long time due to mounts/dismounts of TONS of volumes... but that might  be
only once or twice a year... other than that, all our DB's are complete
archives each time so all their data that they restore is on a select few
tapes and collocation wouldn't buy anything anyway.

OK, I'm reaching the end of a page so I'll cut my reply here ;-)

later,
        Dwight