An attendee within the Constructing Subsequent-Technology Knowledge Heart on-line course despatched me an fascinating dilemma:
Some clients don’t like EVPN due to complexity (it’s required information BGP, symmetric/uneven IRB, ARP suppression, VRF, RT/RD, and so on). They agree, that EVPN offers extra stability and broadcast site visitors optimization, however nonetheless, it won’t save DC from broadcast storms, as a result of protections strategies are the identical for each options (decrease L2 segments, storm-control).
We’ll cope with the pointless EVPN-induced complexity another time, at the moment let’s begin with a couple of intro-level particulars.
The traditional means of constructing knowledge middle materials was to deploy MLAG clusters on the leaf- and the backbone layer, pretending the backbone layer is a single “node”, and working STP to forestall any potential forwarding loops. Core convergence relied on LACP, UDLD, and STP.
The currently-hip knowledge middle material design begins with an IP community, provides VXLAN transport on prime of that, and makes use of EVPN because the management airplane. Core convergence depends on BFD, IP routing protocols and Quick Reroute (if wanted).
Earlier than digging into the main points, it’s price noting that it’s completely doable to construct VXLAN-based materials with out EVPN and the related complexity. All it’s a must to do is to configure static ingress replication lists, buying and selling protocol complexity for configuration complexity.
I’m not saying that’s the advisable means of doing issues, nevertheless it’s a viable choice often not talked about by the networking distributors. Simply take into account that EVPN and related complexity is not a compulsory little bit of VXLAN-based materials.
I’d additionally strongly suggest producing the ingress replication lists with an automation instrument if you wish to keep sane.
Which Design Ought to I Use?
I wouldn’t take into consideration constructing bridged material in 2022. MLAG stays a kludge and I’ve seen too many knowledge middle meltdowns brought on by MLAG bugs.
Moreover, constructing a bridged material forces you to make use of MLAG on the backbone layer (the place bugs matter most), even when the end-hosts don’t want hyperlink aggregation – a extremely debatable matter we lined within the December 2021 session of the Design Clinic.
With out going into the main points: you MIGHT want hyperlink aggregation on storage nodes and also you PROBABLY SHOULD NOT use hyperlink aggregation on hypervisor hosts like VMware ESXi. Do you have to agree with this finest apply you can construct a VXLAN-based material with out ever opening the Pandora field of MLAG complexity.
For extra particulars, watch the Leaf-and-Backbone Material Architectures and EVPN Technical Deep Dive webinars.
Need to kick the tires of the EVPN/VXLAN “magnificence”? netlab launch 1.3 added VXLAN and EVPN help for a half-dozen platforms.