Re: A tunneling proposal

[Ramakrishna Gummadi <ramki@cs.berkeley.edu>]

> At 06:20 PM 7/16/01, Ramakrishna Gummadi wrote:
> >One can think of a simple extension to tunneling under such ISP-wide
> >outages---if tunnel creation fails, the second ISP initiates
> >non-aggregatable route announcements for the prefixes from the address
> >space of the first ISP who has become unreachable. This two step approach
> >may prove effective for fixing both small-scale temporary and large-scale
> >persistent problems.
>
> This could work, provided the alternate upstream ISPs are willing, and
> other networks are willing to take the announcements. The concern is that
> RADB and similar are used by providers to restrict where announcements will
> come from. Keep in mind also that what we're talking about here is human
> intervention in the case of an outage. The question to be answered by
> prospective users of such a multhoming solution is how many hours of outage
> can be tolerated? I suspect it unlikely that automated tunnel creation and
> announcement setup would be well received by those running backbones, so
> there's likely to be a human element at play.

Is it not possible to create a *static* entry in the RADB or other policy
database permitting each of a site's ISPs to advertise the site's prefixes
drawn from other the address space of other ISPs? I understand it is not
straightforward, but because it is static, perhaps it can be automated...

> >It may be a good idea to examine a breakdown of the kind of failures
> >experiences within an ISP, because this will help judge the effectiveness
> >of the tunneling approach. But the primary motivation in tunneling is not
> >to introduce unnecessray routes into the DFZ if there is an alternative.
> >For example, I know that, in the context of server farms, about 20% of
> >outages and errors are because of local power failures, but this may not
> >hold in the ISP case. Does anyone have any info about the main problems
> >and their frequencies?
>
> Where's that 20% figure come from? With all the backups the colo vendor
> salesfolk talk about, that's a surprising figure.

I heard this number from Jim Gray's seminar talk at UC Berkeley when he
was analyzing systems (databases, in particular) failures. But this was
from a couple of years ago....