Data Center
Network Evolution
The Data Center
fabric journey evolved from a Spanning-Tree (STP) based network to more
efficient ways of using the available resources of the infrastructure.
Initially, we built
Spanning-Tree (STP) based networks where a single link was active for a given
service (VLAN). The same was true for the demarcation between Layer 2 and Layer
3, where a First Hop Redundancy Protocol like HSRP or VRRP provided a single
active point of egress per service (VLAN/subnet).
With the
introduction of Cisco’s version of Multi-Chassis Link-Aggregation, virtual
Port-Channels (vPC), a significant improvement has been achieved by providing a
loop-free topology. In vPC-based environments, Spanning-Tree (STP) was still
present to provide a failsafe mechanism without the disadvantages of a single
path tree active. vPC also provided enhancements to the Layer 2 / Layer 3
demarcation, where now First Hop Redundancy Protocols (FHRP) like HSRP and VRRP
start forwarding in an active/active manner, and this improved a former
chokepoint. Nevertheless, even with vPC and FHRP active/active, further
scale-out with regards to Layer 2 and Layer 3 forwarding became an ask.
Cisco introduced
Layer 2 Multipathing (L2MP) to accommodate these asks with the introduction of
Cisco FabricPath. With the MAC-in-MAC frame encapsulation and the IS-IS routing
protocol, Cisco provided a Layer 2 Equal Cost Multipath (ECMP) based network,
where hosts were allowed to talk to other hosts across all available links.
Different to vPC, FabricPath did not require a pairing of network nodes and the
configuration became simplified. A Layer 2 fabric was made available and the
first scale-out network architecture was embraced; the need for a wider Layer 2
/ Layer 3 demarcation became eminent. With Anycast-HSRP, Cisco implemented a
way to scale-out the common chokepoint for First-Hop gateways and extended it
to 4 active nodes. Further scale-out was an ask for FabricPath and enhancements
have been added like Enhanced Forwarding, Distributed Anycast Gateway at the
Leaf switch, and automation of connected workloads; these enhancements fell
under Cisco’s Dynamic Fabric Automation (DFA) solution.
With the industry
moving from Frame Encapsulation (MAC-in-MAC) to Packet Encapsulation
(MAC-in-IP), Cisco embraced VXLAN within its Data Center Switching portfolio to
provide a standards based encapsulation technique. Initially, VXLAN was
introduced as a Layer 2 service only and since VXLAN Flood and Learn (as
defined in RFC 7348) follows the similar Flood and Learn semantic as Ethernet
or FabricPath, enhancements were required. VXLAN with a control plane became
necessary to introduce Layer 2 and Layer 3 services, while optimizing
forwarding to address the limitations of VXLAN Flood and Learn. Multiprotocol
Border Gateway Protocol with Ethernet Virtual Private Network (MP-BGP EVPN) was
introduced as the control plane, with VXLAN being used in the data plane.
MP-BGP EVPN has been defined by IETF as the standards-based control plane for
VXLAN overlays. The Programmable Fabric solution is based on VXLAN with BGP
EVPN, with the programmability of the network fabric through APIs.
Data Center fabric
challenges remained present and included disjoint provisioning, CLI centric
box-by-box configuration, disruptive growth of applications (and other data
center entities), deficient host overlay, and location dependency (rigid
coupling of IP address to location). These challenges have led to operational
complexity, architectural rigidness and infrastructure inefficiency.
The Programmable
Fabric solution addresses the above challenges by providing the following:
This results in
simplifying of the underlying fabric and optimizing the overlay for north-south
as well as east-west traffic flows, which eases the placement of workload as
the network scales with integrated automation
Feedback