On Wed, 3 Sep 2003, Michael Thomas wrote:
Michael Richardson writes:
I despair that it will not be till way too late
that we discover -- unsurprisingly -- that flash
crowds and IP brownouts are not only possible, but
to be expected. More's the pity that we have both
the standards and the deployed code (RSVP) to largely
avoid this disaster in the making.
This overstates things by quite a bit. RSVP is not up to the task in a
very significant way. Not only is RSVP not scalable (in similar ways that
affect Multicast), but even assuming those problems were solved, which is
possible, even given an ideal labatory setting, RSVP does't solve the
problem,
RSVP is not "too heavy". It doesn't work. Not even a little bit.
Assuming every router in the path played RSVP, there are two show-stopper
problems: It can't detect a network failure, and the path might change for
other reasons, so traffic is flowing places where resources weren't
reserved.
For RSVP to work, every router, end to end has to use it. If ISPs used
RSVP between ISPs, some provider somewhere else would be able to really
screw up your network. (Just like multicast routing)
RSVP was a good try, but not a production solution.
I think that there's some belief that this is a
technical problem with RSVP (too heavy, too
whatever), but I think that's a gloss: signaled
QoS is just plain hard and heavy and slow and all
kinds of other unpleasant things by its very
nature. No technical solution is going to make it
into diffserv or best effort. So we're going to
have a disaster and then Congress will do what the
industry couldn't...
Mike