Re: [armd] Multicast in the data center [was Re: address resolution requirement from hosts to overlay edge nodes. Any opinion?]

[AshwoodsmithPeter <[e-mail]>]

> My impression is that many data centers do not enable IP multicast on
> their routers. That means you can use link-local multicast (which
> works fine within one IP subnet and doesn't really have scaling
> issues). But if you want multicast that goes beyond one link (and IP
> subnet), which is presumably necessary for an overlay like
> VXLAN/NVGRE, that is where you have problems.

Agreed. If you extrapolate a bit and imagine VXLAN with 1000's of logical groups.. then that maps either to head end replication in the hypervisor, or to IGMP from the hypervisor and PIM between the TOR and CORE 'switches' and the gateway router.

> The question is not even whether L3 multicast scales. It's whether the
> DC operater is willing to enable such multicast.

No choice if you want to run a Layer 2 Emulation over IP and you care about multicast banwdidh effeciency.

This highlights one of the major differences between the L2 over IP and L2 over L2 solutions, i.e. signalling v.s. computation for the multicast state.

This is why I think an understanding of the true uses of that multicast is important because if 90% of the applications are only finding each other and signalling each other using multicast there are better ways but if they are truly multicasting data in large volumes .. well we need extremely good underlay multicast scale.

Peter