SRv6

Wed Sep 16 23:30:19 UTC 2020

Mark,

> On 16 Sep 2020, at 10:32, Mark Tinka <mark.tinka at seacom.com> wrote:
> 
> On 15/Sep/20 19:00, aaron1 at gvtc.com wrote:
> 
>> Sorry guys, I'm not aware of much of what you mention as far as agenda, vendor motive, and hardware support, etc.... 
> 
> I'm not shy... this would be Cisco.

And that’s fine. The fact that some Intellectual Property[1] was created
by one vendor or another (disclaimer - I work for Cisco) shouldn’t be
by default something that discredits the idea. And BTW, if You look into
the history of how SR started, it was close feedback loop with at least
some of the ISPs wanting to have “easier” and SDN-ish control over the
network so the blame should be shared :) Having support from other
vendors was de facto requirement to even think about deploying it widely
and that's better approach IMHO than “lets patent everything out and
force everyone into our black-box-architecture that’s best in the world”.

I’m observing the discussions over the last couple of months and
generally they boil down to “leave us alone, everything sucks, we’ll 
stay with what we have”. And sure, as you indisputably proven during last
30 years, running modern ISP network can be done over IP, MPLS, and the
network can even survive introduction of IPv6. And I get it - vendors
have generally failed to address your ideas and problems in timely manner,
and when we did, quality was simply not there. But really, is that all
what’s interesting in life? I doubt it. Unless the whole point of
discussing here would be to start from technical topics
(because of ‘rules’) and end up with everybodys favorite part - beating
virtual Pinata made to look like representation of most hated
salesman/vendor. Then sorry, I somehow missed that :)

While I personally find the idea of stacking IPv6 labels gruesome for
any non-trivial label depth[2] (and I'm really worried about software guys
coming in from the “mobile app” world soon, and finding out that they can
create those IPv6 EH stacks easily), going forward we have to think
about what will keep networks running in for the next 5, 10 or 20 years.
IPv4 with MPLS+LDP+RSVP-TE will work fine of course, so will SRoMPLS,
but IPv6 is gaining adoption and need to multiplex/demultiplex more and
more services and users will only grow. And BTW, MPLS forwarding between
ASes in the internet is still something that works mainly on slides,
highly paid consulting “proposals” and of course on the CCIE exam.

On the other side, there’s Elon Musk moving us to Mars, wretched IoT
world with “build, sell and forget" mentality w/r to firmware and good
network stacks. And yet only 59% people around the world today have internet
access. At least good portion of that heavily filtered one by the way.

IPv6 seems to be good plan forward (and would potentially unify
architecture of normal routing and overlay routing with SRv6), even
if things like MPLSoUDP or GRE would really solve everything if pushed
with enough force[3] ;) It’s worth observing, that from this perspective
IPinIP would be as good as SRv6 if everyone would agree 20 years ago that
source routing is acceptable. LDP or RSVP-TE would never gain any usage
and maybe we wouldn’t learn lesson or two. BTW, we tried to somehow get
back to this simplification with LISP[4], and in the long run it seems
overloading address semantics is not something that is happily accepted
everywhere, and it doesn’t matter if that's IPv4 or IPv6.

So much hated middleboxes will also adopt faster to IPv6, as adopting MPLS
data plane after those 20 years on firewalls, load balancers and what-you
looks kind of dissapointingly. And if we are talking about network
functions - I believe it’s more important right now to agree on one way
of doing service chaining, than discussing SR or SRv6 as evil seed created
to conquer the world. 

SR takes out state out, and SRv6 has the same address format on the
outside as well as inside. You can happily run it with both data planes,
and while today maybe you can’t provide migration of ALL services,
SR+IGP quite nicely interworks with MPLS+LDP.

Will HW evolve? It has to anyway, no previous change was done day one
and 128 bits times 5 or 8 or 12 seems horrible only today. Over the
years, people got used to bigger horrors ;)

— 
./

[1]. https://patents.google.com/patent/US20140169370
[2]. Yeah, Binding SIDs of course, but that’s a solution to self-made
     problem as Ivan Pepelnjak would say.
[3]. https://tools.ietf.org/html/rfc1925 point 2.3.
[4]. https://tools.ietf.org/html/rfc6830