IPv6 dual stacking and route tables

Mailman List Mirror (Read Only)
1

So, we are preparing to add IPv6 to our multi-homed (separate routers and carriers with IBGP) multi-site business. Starting off with a lab of course. Circuits and hardware are a few months away. I'm doing the initial designs and having some delivery questions with the carrier(s). One interesting question came up. There was a thread I found (and have since lost) regarding what routes to accept. Currently, in IPv4, we accept a default route only from both carriers at both sites. Works fine. Optimal? No. Significantly negative impact? No. In IPv6, I have heard some folks say that in a multi-homed environment it is better to get the full IPv6 table fed into both of your edge routers. Ok. Fine. Then, The thread I was referring to said that it is also advisable to have the entire IPv4 table fed in parallel. Ok. I understand we are talking about completely separate protocols. So it's not a layer 3 issue. The reasoning was that DNS could potentially introduce some latency.

"If you have a specific route to a AAAA record but a less specific route to an A record the potential is for the trip to take longer."

That was the premise of the thread. I swear I googled it for 20 minutes to link before giving up. Anyway, can anyone who's been thru this provide any opinions on why or why not it is important to accept the full IPv6 table AND the full IPv4 table? I have the hardware to handle it I'm just not sure long term what the reasoning would be for or against. Again, I'm an end customer. Not a carrier. So my concern is (A) my Internet facing applications and (B) my users who eventually will surf IPv6.

Any guidance would be appreciated. Thanks.

-Hammer-

"I was a normal American nerd"
-Jack Herer

2

We have been accepting our upstreams' connected and customer routes only (v4) and full routes (v6) without issue. I can't say that I have previously heard of the DNS performance example/concern you provided above

3

You should accept the full v6 table, because some IPs may not,
currently, be reachable via one of the carriers.

4

We currently take full v4 and v6 routes, however we do not yet have end-users officially on v6 (users doing their own 6to4 tunnels and stuff like Teredo notwithstanding), so I don't have any experience with the A/AAAA resolution asymmetry you're describing.

jms

5

So, we are preparing to add IPv6 to our multi-homed (separate routers and
carriers with IBGP) multi-site business. Starting off with a lab of course.
Circuits and hardware are a few months away. I'm doing the initial designs
and having some delivery questions with the carrier(s). One interesting
question came up. There was a thread I found (and have since lost) regarding
what routes to accept. Currently, in IPv4, we accept a default route only
from both carriers at both sites. Works fine. Optimal? No. Significantly
negative impact? No. In IPv6, I have heard some folks say that in a
multi-homed environment it is better to get the full IPv6 table fed into
both of your edge routers. Ok. Fine. Then, The thread I was referring to
said that it is also advisable to have the entire IPv4 table fed in
parallel. Ok. I understand we are talking about completely separate
protocols. So it's not a layer 3 issue. The reasoning was that DNS could
potentially introduce some latency.

"If you have a specific route to a AAAA record but a less specific route to
an A record the potential is for the trip to take longer."

That was the premise of the thread. I swear I googled it for 20 minutes to
link before giving up. Anyway, can anyone who's been thru this provide any
opinions on why or why not it is important to accept the full IPv6 table AND
the full IPv4 table? I have the hardware to handle it I'm just not sure long
term what the reasoning would be for or against. Again, I'm an end customer.
Not a carrier. So my concern is (A) my Internet facing applications and (B)
my users who eventually will surf IPv6.

Any guidance would be appreciated. Thanks.

Well. I don't really follow the above text.

But, the principle is the same for IPv4 or IPv6. If you take a full
or partial table, your router can make a better choice than just
getting default (only maybe, BGP is never guaranteeing anything about
performance).

That said, in v6, it is a little bit more important, IMHO, to take the
ISP routes instead of just a default since the v6 peering is not as
robust out on the Internet. There are still turf wars going on or
some SPs are still not peering IPv6 in all the places they peer for
IPv4. Less peering = longer latency. But, this situation has improved
dramatically in the last 12 months.

In the end, my guidance is to take "provider routes" or "customer
route" + default. This helps your router make an educated guess
without absorbing all the churn and gunk that a full BGP feed hits
your router with. Make the SP trim those routes on their side so you
don't see it.

CB

6

So, we are preparing to add IPv6 to our multi-homed (separate routers
and carriers with IBGP) multi-site business. Starting off with a lab of
course.

Dear "Hammer",

Welcome to the 21th century. 2012 is going to "the year" (they claim,
again :wink: of IPv6 thus it is better to start before the big launch event
that is this year, (of which there was also one back in in 2004:
http://www.global-ipv6.org/ for the folks who have IPv6 for some time
now :wink: Better late than never some would say.

Circuits and hardware are a few months away. I'm doing the
initial designs and having some delivery questions with the carrier(s).
One interesting question came up. There was a thread I found (and have
since lost) regarding what routes to accept. Currently, in IPv4, we
accept a default route only from both carriers at both sites. Works
fine. Optimal? No. Significantly negative impact? No. In IPv6, I have
heard some folks say that in a multi-homed environment it is better to
get the full IPv6 table fed into both of your edge routers. Ok. Fine.
Then, The thread I was referring to said that it is also advisable to
have the entire IPv4 table fed in parallel. Ok. I understand we are
talking about completely separate protocols. So it's not a layer 3
issue.

The only advantage of more routes, in both IPv4 and IPv6 is that the
path selection can be better. An end-host does not make this decision
where packets flow, thus having a full route or not should not matter
much, except that the route might be more optimal. No DNS involvement
here yet.

The trick comes with Happy-Eyeballs alike setups (especially Mac OSX
Lion and up which does not follow that spec and in which it cannot be
turned off either, which is awesome when you are debugging things...)

Due to HE (Happy-Eyeballs) setups, which can differ per OS and per tool.

Chrome has it's own HE implementation, thus if you run Chrome on a Mac
you will have sometimes one sometimes another connect depending on if
you are using Safari or Chrome for instance as Safari does use the
system provided things. Ping will pick another again etc. It will be
quite random all the time.

The fun with the Mac OS X implementation is that it keeps a local cache
of per-destination latency/speed information.

If you thus have two default routes, be that IPv4 or IPv6, and your
routers are swapping paths between either and thus don't have a stable
outgoing path those latencies will vary and thus the pretty HE
algorithms will be very confused.

This is likely why your "source" recommended to have a full path, as per
sub-prefix the path will become more stable and established than if you
are doing hot-potato with two defaults.

Greets,
Jeroen

7

Thanks Jeroen (and Ryan/Philip/Cameron/Justin/Etc.) for all the online and offline responses. That was fast. The struggle is that I'm having trouble seeing how/why it would matter other than potential latency on the IPv4 side. IPv6 conversations usually involve taking the full table when dealing with multi-homed/multi-site setups. IPv4 I didn't really consider (taking the full table) until I mentioned this to some of my vendors technical folks and it caused a lot of reflection. Not on the L3 part. Just on the DNS part. This appears to be a tough subject area when it comes to V4/V6 deployment strategies. The bottom line is that I'll take whatever the carrier feeds in V6. Just trying to see where I would be missing out by not doing the same in V4. Again, I have the hardware to support it and I really have no reason not to do it. I just want to be able to justify to myself why I'm doing it.

A lot of kinks to work out this year.

-Hammer-

"I was a normal American nerd"
-Jack Herer

8

Why you want non-defaults in both IPv4 and IPv6:
- more possible paths
- less chances of blackholes.

And of course, those paths will be more stable and you don't get
hot-potato swapping between two defaults.

And that in turn allows the Happy Eyeballs mechanisms to do their jobs
much better as they keep a history per host or prefix, they assume IPv6
/48's and IPv4 /24's from what I have seen, in some cases.

Greets,
Jeroen

9

OK. Looking forward to getting the lab up. Since I can handle the volume I'll take both tables. At least in the lab. Looking forward to doing some experiments with DNS just to see what all the fuss is about. Looks like I'll need to order a Mac for the lab. No harm there. :slight_smile:

-Hammer-

"I was a normal American nerd"
-Jack Herer

10

You should accept the full v6 table, because some IPs may not,
currently, be reachable via one of the carriers.

Definitely agreed here, and this is why we take full v6 tables. Especially since one of our upstreams does not peer with at least one other large network; if we took just a default from them, we would likely be sending them traffic which they in turn do not have a route for whereas the other upstream of ours does.

11

On a purely theoretical level, mores specific routes are always better than default.

So, on a purely theoretical level:

IDEAL: Full routes, both protocols
  Advantage: Most information available, theoretically best decisions possible.
  Disadvantage: Router cost rivals national debt of third-world country.

Second best: Full routes IPv6, default or partial routes IPv4
  Advantage: Lots of information available, theoretically best IPv6 decisions.
  Disadvantage: IPv6 might outperform IPv4 (not really a problem in most cases)
  Bigger disadvantage: Some IPv4 destinations might get blackholed from time to time.

Third choice: Default both protocols
  Advantage: At least you're still dual-stacked.
  Disadvantage: All the disadvantages applied to IPv4 above now apply to IPv6, too.

Worst choice: Don't implement IPv6
  Advantage: Absolutely none.
  Disadvantage: You can reach progressively less and less of the internet over time.

However, the real answer is more complex than that. Sometimes the route that looks the best in BGP might not actually be the best and so in some cases with full tables you might send to the provider with the less effective path even though default would have had you going via the more effective path. These circumstances are rare, however, but, BGP has lots of knobs and depending on how well you and your upstreams set those knobs, your experience can be radically better or worse as a result.

If your trip to the A destination via default takes longer than your trip to the AAAA destination via specifics, I'm not seeing a problem. Clients that have IPv6 capability will get a better user experience and clients that don't have IPv6 will get the same experience they got with default-based IPv4 prior to you implementing IPv6.

Owen