Re: draft-ietf-v6ops-addr-select-ps-02

[Florian Weimer <fw@deneb.enyo.de>]

* Brian E. Carpenter:

> On 2008-01-28 00:26, Florian Weimer wrote:
>> I've also noticed that the draft fails to mention that Rule 9, when
>> applied by most of the client population on the Internet, results in
>> worse performance than no destination address sorting at all.
>> 
>> Rule 9 might offer a significant advantage in private deployments which
>> use non-IANA address allocation.  In other cases, routing is typically
>> not hierarchical at all, so that the length of the matching address
>> prefix is meaningless.
>
> Not true if the sites concerned are both using PA space and
> happen to have the same ISP.

Only if they are within the same PA prefix.  And even then, it's not
clear if this is the best choice for the client (or the server,
off-loading to other networks is no longer possible).

> We still have hopes that PA space will be predominant in IPv6, despite
> the PI heresy.

There's too little experience with IPv6 to be sure.  But I wouldn't be
surprised with round-robin is superior there, too.  With round-robin,
it's possible to add more instances, and they take a share of the load.

And it's a further obstacle to renumbering -- what do you do if the new
prefix has got an unfortunate bit pattern, shifting the loads in
unwanted way?

> Can you describe cases in which this rule is actively harmful?

pool.ntp.org and some Debian mirrors experienced rather asymmetric
traffic loads after getaddrinfo() with Rule 9 support had been rolled
out.