Hi Dino. > > I suspect we need to dive a bit deeper into this. i.e., what does > > "dual attached" mean?
> A bare-metal server has 2 physical connections to different boxes > (where a box is a swtich or a router). In some cases, the 2 > connections are active/active, but mostly active/backup. There is > some assumption from the server OSes that the upstream boxes are > layer-2 switches. But it doesn't have to be the case, they can be > layer-3 routers. I think there are multiple scenarios here to consider, and the details matter. Scenario 1: Bare metal server has 2 separate connections to two different boxes. How are these two connections presented to the (bare-metal) TS? 1A) two different interfaces are visible to the TS, both with their own separate IP addresses. This is not an issue for NVO3, since this is modeled as two interfaces to two separate VNs. (Right?) B) TS sees only one interface, the CNA hides the details of two physical uplinks from TS. Links could be active/active or active/backup. In this case, the TS has one IP adddress (I presume) with the CNA hiding the details of multiple uplinks from the TS. This is a messier case... Do both uplinks have to connect to the same NVE? In the standard L2 case, don't both L2 uplinks have to connect to the "same L2 network"? That only means that both ToRs would need to support the same VLANs and connect to the same LAN domains. (Right?) If above we allow the two uplinks to go to two different NVEs, the implication is that from an NVO3 perspective, a given TS mapping is reachable through different NVE addresses. That raises all sorts of interesting questions. :-) > > Do you mean the NVE has two physical connections to the ToR? If so, > No, that would be the NVE being able to get ECMP support from the > underlay. I am talking about the NVE in the TOR, as one example. Right, NVE is doing IP and each uplink would be to different L3 dest. > > than the question presumably is about allowing/supporting multi-homed > > NVEs. > That is right Thomas. Yikes. I guess we need to have that discussion. :) > > If the question is about the TS having two attachements to the ToR, > > wouldn't that be modeled (architecturally) as two distinct > > interfaces/ports, in which case each one can connect to its own NVE, > > in which case we are good? > The questions is about the return traffic coming to the TS. Do you > want it load-split across all the NVEs associated with the TS. The > answer is a definite yes, as Joel stated for robustness and to take > advantage of all cross-sectional cheap bandwidth in a data center. Fundamentally, it means that address mappings (inner to outer) are not one-to-one, but are one-to-many. Is this something we need to support? I agree it has some benefits. But it also adds some complications... Thomas _______________________________________________ nvo3 mailing list [email protected] https://www.ietf.org/mailman/listinfo/nvo3
