Re: [dc] Requirement for a method to manage mac address in DC

[Donald Eastlake <[e-mail]>]

Hi Pat,

On Fri, Feb 3, 2012 at 1:51 PM, Pat Thaler <[e-mail]> wrote:
> Donald,
>
> Everything has to be somewhere.

"remember: no matter where you go... there you are" The Adventures of
Buckaroo Banzai Across the 8th Dimension...

> By your argument, TRILL bridges aren't layer 3 devices either
> because they aren't peering with the routers in your slides. They
> are below layer 3 -

Yes. Exactly the same form of proof that proves that TRILL switches
are above layer 2 also proves that TRILL switches are below layer 3.

> not working in the layer 3 addressing domain so they are in layer 2.

No. TRILL Data packets are routed by TRILL switches using the TRILL
nickname address space. I do not know what the basis is of your
assertion that TRILL nicknames are not layer 3 addresses but I don't
believe they are layer 2 addresses.

> It is just that layer 2 has some sublayers of peered devices.
> Long before TRILL and PBB, there were Ethernet repeaters which were
> also layer 2 devices and didn't peer with switches. Provider
> bridges, provider backbone bridges and TRILL all work at different
> sublayers in layer 2.

I disagree the TRILL is a sublayer of Layer 2. It is fairly easy to
order devices as to relative layer based on peering although, like
with anything else, if you apply a sufficiently strong magnifying
glass you can find some odd glitches and corner case:

Repeater < Prov. Bridge < Cust. Bridge < TRILL Switch < L3 Router

In my opinion, the arguments that TRILL is in Layer 2 or Layer 3 are
in exact equipoise. I do not agree that anything on or inside the
border betwen Layer 2 and Layer 3 should be classified as Layer 2.

> I also don't see how that matters to the content of this
> discussion. Whether one considers TRILL bridges to be at some new

TRILL routers

> layer 2.5 or not or even at layer 3 doesn't matter to the point that
> they provide isolation between the address spaces of tenants. That
> isolation only applies if any traffic between those tenants goes
> through a layer 3 devices that removes the original MAC addresses
> from the frame.

I agree that it does not matter much to the point under discussion. I
just didn't want people to be confused about the true nature of TRILL
switches.

Thanks,
Donald
=============================
 Donald E. Eastlake 3rd   [phone] (cell)
 155 Beaver Street, Milford, MA 01757 USA
 [e-mail]

> Regards,
> Pat
>
> -----Original Message-----
> From: [e-mail] [mailto:[e-mail]] On Behalf Of Donald Eastlake
> Sent: Friday, February 03, 2012 10:18 AM
> To: Pat Thaler
> Cc: Thomas Narten; [e-mail]; yu jinghai; Truman Boyes; Lizhong Jin; Mallik Mahalingam
> Subject: Re: [dc] Requirement for a method to manage mac address in DC
>
> Hi Pat,
>
> Please see below:
>
> On Thu, Feb 2, 2012 at 3:01 PM, Pat Thaler <[e-mail]> wrote:
>> Some work on managing MAC addresses of virtual devices in a Data Center may
>> be worthwhile, though it isn't clear to me whether such work would better
>> fit in IETF or IEEE 802.
>>
>>
>>
>> When virtualization ecosystem management entities are handing out addresses,
>> there can be data centers with multiple such entities and one can't count on
>> them to coordinate their use of the address space. While each of them won't
>> hand out duplicate addresses to the set of VMs they manage, the addresses
>> may be duplicated for VMs managed by different management entities.
>> Sometimes this can be dealt with by manual assignment of ranges, but in a
>> data center with multiple tenants, the tenants are unlikely to coordinate
>> that. The potential duplicate addresses can in some cases be dealt with by
>> mechanisms that keep the address space of the management entities separate
>> such as: IVL (or other mechanisms that concatenate VLAN and MAC address for
>> bridge learning) or layer 2 (e.g. PBB and TRILL) or layer 3 encapsulations.
>
> Sorry to be nit-picky, but TRILL is not a layer 2 encapsulation. It is
> provably above layer 2.
>
> In my opinion, the best way to tell if a device of type X is at a
> higher layer, at the same layer, or at a lower layer, than a device of
> type Y is to look at peering. Generally speaking, layer 2 devices are
> transparent to TRILL and TRILL switches peer through layer 2 devices,
> just like layer 3 routers peer with each other through layer 2
> devices. On the other hand, TRILL switches look like end stations to
> and block peering between layer 2 devices, just like layer 3 routers
> look like end stations and block peering between layer 2 devices. See
> attached slides.
>
> Thanks,
> Donald
> =============================
>  Donald E. Eastlake 3rd   [phone] (cell)
>  155 Beaver Street, Milford, MA 01757 USA
>  [e-mail]
>
>>  But there could be some areas where a protocol for coordinating assignments
>> to avoid duplication would help.
>>
>> There have been discussions in the IEEE RAC about concerns regarding the use
>> of MAC addresses from the global MAC address space for virtual devices;
>> issues include potential for exhausting the global address space and that an
>> address that looks like a global address is being used as a local address.
>> Half the MAC address space is for local addresses, but there aren't
>> standardized mechanisms for managing addresses in that space.
>>
>>
>>
>> <IEEE 802 Vice-Chair hat on> If work was done in the IETF on MAC address
>> management/assignment, there should be close liaison with IEEE 802 and the
>> IEEE RAC.
>>
>>
>>
>> Pat
>>
>>
>>
>> From: [e-mail] [mailto:[e-mail]] On Behalf Of Mallik
>> Mahalingam
>> Sent: Thursday, February 02, 2012 11:22 AM
>> To: Truman Boyes
>> Cc: Thomas Narten; yu jinghai; [e-mail]; Lizhong Jin
>>
>>
>> Subject: Re: [dc] Requirement for a method to manage mac address in DC
>>
>>
>>
>> In a virtualized environment MAC addresses are not totally random generated.
>> There is some notion of Management-Entity(s)/controller(s) allocating the
>> MAC addresses for VMs and ensures that it does not assign the same MAC
>> address to two different VMs and this work only within the scope of that
>> management/controller administration. There are some exceptions of course
>> (a) MAC address exhaustion under a given OUI category  (b) manual
>> copy/cloning of VMs and powering on them using standalone management
>> entities (c) VMs that use MAC address override for legitimate reasons
>> [because else things like licensing software breaks].  There are some
>> mechanisms in place to address (a), but (b) and (c) requires co-operation at
>> the management-entity/controllers.
>>
>> Mallik
>>
>> ________________________________
>>
>> From: "Truman Boyes" <[e-mail]>
>> To: "Thomas Narten" <[e-mail]>
>> Cc: "yu jinghai" <[e-mail]>, [e-mail], "Lizhong Jin"
>> <[e-mail]>
>> Sent: Thursday, February 2, 2012 10:20:07 AM
>> Subject: Re: [dc] Requirement for a method to manage mac address in DC
>>
>>
>> On Thu, Feb 2, 2012 at 10:55 AM, Thomas Narten <[e-mail]> wrote:
>>
>> Truman Boyes <[e-mail]> writes:
>>
>>> The L2 separation between multiple tenants is true in most circumstances
>>> in
>>> DCs, but in commodity computing (ie. VPS, low cost dedicated servers, or
>>> co-location) there is a concern on IPv4 address exhaustion or waste, so
>>> machines/instances are grouped on single L2 segments. It is possible to
>>> have virtual MAC overlaps on these segments. Is this something that this
>>> group wishes to evaluate options to solve?
>>
>> IMO, this is putting the cart before the horse.
>>
>> Can we first get a sense for how big a problem this is in practice and
>> whether existing mitigation approaches are not sufficient?
>>
>> I.e., is this a real problem causing significant pain today, or are
>> their other bigger "pain points" that we should be looking at?
>>
>> Thomas
>>
>>
>> In the VPS/VM world,  I would say it's not a significant issue because there
>> are single entities (Organizations) that manage the MAC addresses. Typically
>> software would just increment the virtual MACs, and this does not require
>> external protocols to ensure uniqueness. If there are many provisioning
>> systems that manage VMs on the same network segment then they will need to
>> keep their database in sync.
>>
>> --
>> --truman