This week Cisco announced what amounts to a total re-work to their existing FabricPath virtualized multi-tenant data center design. The marketing name given to this DFA – Dynamic Fabric Architecture, aka – Vinci. The aim is to accomplish a dynamically configured data center fabric.
Summary: Fundamentally what’s new with DFA is pushing layer 3 to the access-edge/leaf while still allowing any host to reside on any leaf in the fabric. This takes the load off of the spine to have to do any learning, routes or MAC, thus the spine becomes something akin to an MPLS core router. The last major component of DFA is the reworking of DCNM to enable it to be more of an auto-provisioning and fabric orchestration tool.
How it works: If you understand FabricPath and conversational learning then you are well on the way to understanding DFA. As you may expect FabricPath is the core connectivity protocol. However, by using an anycast IP on the leaf plus MP-iBGP (over IS-IS) to carry the learned hosts/prefixes enables layer 3 at the edge. Along with some other mechanisms running in the background such as proxy ARP, this allows any VM/host to be connected to any leaf node.
- Foundation connectivity based on FabricPath.
- Uses POAP (Power On Auto Provisioning) to auto-configure the network fabric.
- New version of DCNM (totally new/different) used to provision and manage infrastructure
- VDP (VSI Discovery & Provisioning) – runs between hypervisor and leaf & is used for auto-provisioning of network configuration for new VMs – part of 802.1Qbg
- Auto-Configuration for physical hosts – switches intercept and forward traffic to DCNM which profile the traffic and push
- Layer 3 gateway on the leaf & in theory, up to 384 anycast HSRP gateways since DFA supports up to 384 leaf switches
- Any/all subnets on any leaf because the leaf nodes are the HSRP
- Spine & leaf is supported but w/ DFA the preferred topology is folded CLOS
- Will still use vPC+ to connect hosts
- iBGP is used as the control plane to distribute internal host/subnet routes & exchange reach-ability information
- ARP is not flooded across the fabric fundamentally via proxy-ARP
- Segment ID: DFA uses a double Dot1q tag for more VLANs, not Q-in-Q tunneling
- OVA – Open Virtualization Appliance
- Open protocols – XMPP, XMl, SNMP, REST, LDAP, VDP, LLDP, AMQP
- 5596s can potentially be re-used as dedicated spines in a DFA design
- Limitation is that you cannot run your BPG RRs on the 5596 spines
- May not be officially supported
- Multicast optimization by pruning multicast tress at the spine for any spine-leaf pairs that have no receivers connected to them