[EMAIL PROTECTED] (Stephane Bortzmeyer) writes:

> On Thu, Sep 30, 2004 at 08:48:37PM -0400,
>  Dean Anderson <[EMAIL PROTECTED]> wrote 
>  a message of 56 lines which said:
> 
> > If Av8 turns on PPLB, traffic to F-root will go through both sprint
> > and att on a per-packet basis.

Um.

I just don't think that's true.  The default BGP configuration on vendor-C
and vendor-J and presumably every other vendor is that only the best path
is copied from the RIB to the FIB for any given prefix.  So, in order for
Av8 (or anyone else) to see per-packet load balancing toward an f-root (or
any other anycasted service), the following conditions would have to obtain:

1. override the default on BGP's movement of routes from the RIB to the FIB;
2. see the same prefix aspath-length and MED from multiple transits/peers;
3. override the default to change per-destination to per-packet LB.

It's worth noting that since this is not an example of link-bundling (for
which PPLB was put into IOS), the demerits noted in UniverCD about PPLB will
apply -- mainly that video and voice quality will be poor.  Also, TCP
performance overall will be lower.

It's also worth noting that from Av8's (or anyone's) perspective, anycast is
indistinguishable from multihoming.  That means that during the years f-root
was connected to 100++ BGP peers at PAIX, the world saw 100++ paths to f-root
and would have had exactly the same problems with PPLB.

Finally, it's worth noting that even if somehow DNS were the only service
affected (and not voip or video streams or etc), and even if somehow a remote
BGP listener were able to distinguish between multihoming and anycast (which
it can't), then only TCP would be injured, and DNS is sigma-6 UDP.

> Troll Bot <[EMAIL PROTECTED]> keeps mentioning PPLB. May be some people
> more knowledgeable about BGP than I am will explain to me why PPLB is
> such a new issue for anycasting?

It isn't.  The troll in question has only negative objectives, and too much
time on his hands, and this thread is just his latest attempt to try to make
enough smoke that somebody somewhere will worry that there's a fire burning.

> Even without PPLB, the simple and normal (though infrequent) change of
> the routes by BGP may disturb existing TCP sessions if the target is
> anycasted.  This is why anycast is currently deployed only on
> mostly-UDP services like the DNS.

Yes.

> So, it seems there is nothing new coming from the PPLB thing.

It's much worse than that, but at a minimum, yes.
-- 
Paul Vixie
.
dnsop resources:_____________________________________________________
web user interface: http://darkwing.uoregon.edu/~llynch/dnsop.html
mhonarc archive: http://darkwing.uoregon.edu/~llynch/dnsop/index.html

Reply via email to