Home > Articles > VXLAN/EVPN Forwarding Characteristics

VXLAN/EVPN Forwarding Characteristics


  1. Multidestination Traffic
  2. VXLAN BGP EVPN Enhancements
  3. Summary

Chapter Description

In this chapter from Building Data Centers with VXLAN BGP EVPN: A Cisco NX-OS Perspective, the author team discuss two different options for handling BUM or multidestination traffic— leveraging multicast replication in the underlying network and using a multicast-less approach called ingress replication.

VXLAN BGP EVPN Enhancements

The following sections describe some of the feature enhancements that ride on top of the BGP EVPN control plane, further enhancing the forwarding of Layer 2 and Layer 3 traffic in a VXLAN fabric.

ARP Suppression

Address Resolution Protocol (ARP) is responsible for IPv4 address–to–MAC address mapping in IP networks. ARP facilitates obtaining endpoint MAC address information, leveraging the IP address information sent out in an ARP broadcast request. The broadcast request is treated as multidestination traffic that is typically VXLAN encapsulated and sent to every VTEP or edge device that is a member of the corresponding Layer 2 VNI. The response to the ARP request is typically sent as a unicast packet to the requestor. ARP traffic is scoped within the bounds of the broadcast domain represented by the Layer 2 VNI. Correspondingly, IPv6 networks rely on the Neighbor Discovery Protocol (ND) for IPv6 address–to–MAC address resolution. With IPv6, this resolution is triggered via an initial neighbor solicitation (NS) that is multicast through the Layer 2 broadcast domain. A neighbor advertisement (NA) sent in response from the destination completes the neighbor resolution.6

When an endpoint needs to resolve the default gateway, which is the exit point of the local IP subnet, it sends out an ARP request to the configured default gateway. The ARP operation allows the local edge device or VTEP to populate IP/MAC mappings of the locally attached endpoints. In addition to the MAC-to-IP table being populated on the edge device, all the MAC information is populated to the BGP EVPN control plane protocol. In addition, for Layer 2, Layer 2 VNI, Route Distinguisher (RD), Route Target (RT), hosting VTEP, and associated IP information is populated, and for Layer 3, Layer 3 VNI, RD, RT, hosting VTEP, and RMAC information is populated. Specifically, this information is populated in a Route type 2 advertisement and distributed to all remote VTEPs, which now have learned about that endpoint.

Typically, all ARP broadcast requests sent out from an endpoint are subject to copy-to-router and forward behavior. The local edge device learns about an endpoint as long as the endpoint sends out some type of ARP request—not necessarily an ARP request for the resolution of the default gateway hosted on the local edge device.

Typically, when an endpoint wants to talk to another endpoint in the same subnet, it sends out an ARP request for determining the IP-to-MAC binding of the destination endpoint. The ARP request is flooded to all the endpoints that are part of that Layer 2 VNI. ARP snooping coupled with the BGP EVPN control plane information can help avoid flooding for known endpoints. By using ARP snooping, all ARP requests from an endpoint are redirected to the locally attached edge device. The edge device then extracts the destination IP address in the ARP payload and determines whether it is a known endpoint. Specifically, a query is done against the known endpoint information from the BGP EVPN control plane.

If the destination is known, the IP-to-MAC binding information is returned. The local edge device then performs an ARP proxy on behalf of the destination endpoint. In other words, it sends out a unicast ARP response toward the requestor with the resolved MAC address of the known destination endpoint. In this way, all ARP requests to known endpoints are terminated at the earliest possible point, which is the locally attached edge device or VTEP or leaf. This is known as ARP suppression. ARP suppression is possible because all the information is known on all the VXLAN VTEPs.7

Note that ARP suppression is different from the Proxy ARP,8 in which the edge device or router may serve as a proxy on behalf of the destination endpoint, using its own Router MAC. ARP suppression reduces ARP broadcast traffic by leveraging the BGP EVPN control plane information. ARP suppression is enabled on a per-Layer 2 VNI basis. In this way, for all known endpoints, ARP requests are sent only between the endpoint and the local edge device/VTEP. Figure 3-4 illustrates a scenario where ARP suppression at VTEP V1 results in early ARP termination of a request for a host from a host in Layer 2 VNI 30001. It is important to note that the ARP suppression feature works based on the knob enabled under the Layer 2 VNI, regardless of whether the default gateway is configured on the leafs.

Figure 3-4

Figure 3-4 ARP Suppression

When the destination endpoint is not known to the BGP EVPN control plane (that is, a silent or undiscovered endpoint), the ARP broadcast needs to be sent across the VXLAN network. Recall that ARP snooping intercepts the ARP broadcast request to the locally attached edge device. Because the lookup for the destination endpoint results in a “miss,” the edge device re-injects the ARP broadcast request back into the network with appropriate source filtering to ensure that it does not return to the source port from which the ARP request was originally received. The ARP broadcast leverages the multidestination traffic forwarding implementation of the VXLAN fabric.

When the queried endpoint responds to the ARP request, the endpoint generates an ARP response. The unicast ARP response is also subjected to ARP snooping behavior so that it is sent to the remote edge device to which the destination is directly attached. Consequently, the IP/MAC binding of the destination endpoint is learned by the remote VTEP, resulting in that information being populated into the BGP EVPN control plane. The remote edge device also sends out the ARP response to the original requesting endpoint, thereby completing the end-to-end ARP discovery process. The ARP response is sent via the data plane to ensure that there are no delays in the control plane updates. It is important to note that after the initial miss, all the subsequent ARP requests for the discovered endpoint are handled with local ARP termination, as described earlier.

Figure 3-5 illustrates a scenario involving ARP termination or ARP suppression where a miss occurs in the control plane. Host A and Host B belong to the same IP subnet, corresponding to the Layer 2 VNI 30001. Host A wants to communicate with Host B, but the respective MAC-to-IP mapping is not known in Host A’s ARP cache. Therefore, Host A sends an ARP request to determine the MAC-to-IP binding of Host B. The ARP request is snooped by VTEP V1. VTEP V1 uses the target IP address information gleaned from the ARP request payload to look up information about Host B in the BGP EVPN control plane. Since Host B is not yet known in the network, the broadcast ARP request is encapsulated into VXLAN with the destination IP address of a multicast group or a unicast destination IP address. The multicast group is leveraged when multicast is enabled in the underlay. Otherwise, using head-end replication, multiple unicast copies are generated to each remote interested VTEP, as mentioned earlier.

Figure 3-5

Figure 3-5 ARP lookup miss with ARP Suppression

The ARP broadcast sent to VTEP V2 and VTEP V3 is appropriately decapsulated, and the inner broadcast payload is forwarded to all Ethernet ports that are members of VNI 30001. In this way, the ARP request reaches Host B, and Host B answers with a unicast ARP response back to Host A. Once the unicast ARP response hits VTEP V2, the information is snooped, and the control plane is then updated with Host B’s IP/MAC address information. At the same time, the ARP response is encapsulated and forwarded across the VXLAN network to VTEP V1. On receiving the ARP response, VTEP V1 decapsulates the VXLAN header and forwards the decapsulated ARP response, based on a Layer 2 lookup, in VNI 30001. Consequently, Host A receives the ARP response and populates its local ARP cache. From this point forward, bidirectional communication between Host A and Host B is enabled because all MAC and IP address information of both hosts is known in the network.

The ARP request and response process between endpoints in a VXLAN BGP EVPN fabric is similar to that in classic Ethernet or any other Flood and Learn (F&L) network. However, the advantage of the control plane and ARP suppression is realized with the BGP EVPN information. A single ARP resolution of an endpoint triggers the control plane to populate that endpoint information across the entire network, at all the VTEPs. Consequently, any subsequent ARP requests to any known endpoint are locally answered instead of being flooded across the entire network.

Because ARP snooping proactively learns the endpoint information to populate the control plane, it certainly helps in reducing unknown unicast traffic. ARP requests really need to be flooded only for the initial period, when an endpoint has yet to be discovered. However, there are other sources of unknown unicast traffic (for example, vanilla Layer 2 non-IP traffic or even regular Layer 2 IP traffic) that may suffer a DMAC lookup miss and the traffic being flooded all across the VXLAN network.

To completely disable any kind of flooding due to unknown unicast traffic toward the VXLAN network, a new feature called unknown unicast suppression has been introduced. This independent feature is enabled on a per-Layer 2 VNI basis. With this knob enabled, when any Layer 2 traffic suffers a DMAC lookup miss, traffic is locally flooded, but there is no flooding over VXLAN. Therefore, if no silent or unknown hosts are present, flooding can potentially be minimized in a BGP EVPN VXLAN network.

Next, we consider the third piece of BUM or multidestination traffic, which is called multicast traffic. Layer 2 multicast in a VXLAN network is treated as broadcast or unknown unicast traffic because the multicast traffic is flooded across the underlay. Layer 2 multicast flooding leverages the configured multidestination traffic-handling method, whether it exists in unicast mode or multicast mode. Not all platforms can differentiate between the different types of Layer 2 floods, such as multicast or unknown unicast. The IGMP snooping feature presents an optimal way of forwarding Layer 2 multicast traffic if supported by the platform. The Cisco NX-OS default configuration results in behavior that floods Layer 2 multicast traffic both locally and across the VTEP interfaces (see Figure 3-6). This means that all participating VTEPs with the same VNI mapping and all endpoints in that Layer 2 VNI receive this Layer 2 multicast traffic. The multidestination traffic is forwarded to every endpoint in the VNI, regardless of whether that endpoint is an interested receiver.

Figure 3-6

Figure 3-6 No IGMP Snooping in VLAN/VXLAN

With VXLAN, the option to implement IGMP snooping in the same way it is implemented in traditional Ethernet-based networks. The main difference is that the VTEP interface selectively allows Layer 2 multicast forwarding for interested receivers that are present behind a remote VTEP. If there are no interested receivers behind remote VTEPs, Layer 2 multicast traffic received from a source endpoint is not sent toward the VXLAN network. However, as long as there is at least one interested receiver behind some remote VTEP, Layer 2 multicast traffic is forwarded to all remote VTEPs since they are all part of the same Layer 2 VNI. The VTEP interface is a multicast router port that participates in IGMP snooping, thereby allowing Layer 2 multicast forwarding. This is dependent on the receipt of the IGMP “join” message from an endpoint for a given multicast group in a Layer 2 VNI.

The received IGMP join report is VXLAN encapsulated and transported to all remote VTEPs that are members of the corresponding Layer 2 VNI. Note that the multicast group(s) employed for overlay multicast traffic between endpoints should not be confused with the underlay multicast group that may be associated with the Layer 2 VNI. Recall that the latter is employed to provide the destination IP address in the outer IP header for forwarding multidestination traffic in the underlay when using multicast mode.

With IGMP snooping enabled for VXLAN VNIs, the ability to realize the true benefits of multicast occurs: Traffic is forwarded only to interested receiver endpoints (see Figure 3-7). In other words, if no interested receivers exist for a given multicast group behind a VTEP in the same VNI, the Layer 2 multicast traffic is silently dropped and not flooded. After decapsulation at the remote VTEP, Layer 2 multicast traffic is forwarded based on the local IGMP snooping state. IGMP join messages are used to selectively enable the forwarding of certain Layer 2 multicast traffic.

Figure 3-7

Figure 3-7 IGMP Snooping in VLAN/VXLAN

Depending on the platform software support, IGMP snooping for VXLAN-enabled VLANs may not be supported. IGMP snooping is a software feature that does not have any dependencies on the underlying hardware. While this section provided a brief overview, detailed IP multicast flows in a VXLAN BGP EVPN network are presented in Chapter 7 “Multicast Forwarding”.

Distributed IP Anycast Gateway

In order for two endpoints in different IP subnets to communicate with each other, a default gateway is required. Traditionally, the default gateway has been implemented centrally at the data center aggregation layer, in a redundant configuration. For an endpoint to reach the centralized default gateway, it has to first traverse a Layer 2 network. The Layer 2 network options include Ethernet, vPC, FabricPath, and even VXLAN F&L. Communication within the same IP subnet is typically bridged without any communication with the centralized default gateway. For routing communication between different IP networks/subnets, the centralized default gateway is reachable over the same Layer 2 network path.

Typically, the network is designed to be redundant as well as highly available. Likewise, the centralized gateway is highly available. First-hop redundancy protocols (FHRP) were designed to support the centralized default gateway with redundancy. Protocols such as HSRP,9 VRRP,10 and GLBP11 are examples of FHRPs. HSRP and VRRP have a single node responsible for responding to ARP requests as well as routing traffic to a different IP network/subnet. When the primary node fails, the FHRP changes the operational state on the backup node to master. A certain amount of time is needed to fail over from one node to another.

FHRPs became more versatile when implemented with virtual PortChannels (vPCs), which allow both nodes to forward routing traffic and permit only the master node to respond to ARP requests. Combining vPC and FHRP significantly enhances resiliency as well as convergence time in case of failures. Enabling ARP synchronization for the FHRP in vPC environments enables ARP synchronization between the vPC primary and the vPC secondary. FabricPath with anycast HSRP12 increases the number of active gateways from two to four nodes. The FHRP protocol exchange and operational state changes are still present with anycast HSRP. With FHRPs, the default gateway, implemented at the data center aggregation layer, provides a centralized default gateway as well as the Layer 2/Layer 3 network boundary.

The increasing importance of Layer 2 and Layer 3 operations, especially in a large data center fabric, has demanded additional resiliency in comparison to what has been provided by traditional FHRP protocols. Moving the Layer 2-Layer 3 network boundary to the fabric leaf/ToR switch or the access layer reduces the failure domain (see Figure 3-8). The scale-out approach of the distributed IP anycast gateway dramatically reduces the network and protocol state. The distributed IP anycast gateway implementation at each fabric leaf no longer requires each endpoint to traverse a large Layer 2 domain to reach the default gateway.

Figure 3-8

Figure 3-8 Gateway Placement

The distributed IP anycast gateway applies the anycast13 network concept “one to the nearest association.” Anycast is a network addressing and routing methodology in which the data traffic from an endpoint is routed topologically to the closest node in a group of gateways that are all identified by the same destination IP address. With the distributed IP anycast gateway (see Figure 3-9), the default gateway is moved closer to the endpoint—specifically to the leaf where each endpoint is physically attached. The anycast gateway is active on each edge device/VTEP across the network fabric, eliminating the requirement to have traditional hello protocols/packets across the network fabric. Consequently, the same gateway for a subnet can exist concurrently at multiple leafs, as needed, without the requirement for any FHRP-like protocols.

Redundant ToRs are reached over a Multi-Chassis Link Aggregation bundle (MC-LAG) technology such as vPC. Port channel hashing selects only one of the two available default gateways, and the “one to the nearest association” rule applies here. The VXLAN BGP EVPN network provides Layer 2 and Layer 3 services, and the default gateway association exists between the local edge device and the endpoint. When the endpoint tries to resolve the default gateway, the locally attached edge device is the only one that traps and resolves that ARP request. In this way, every edge device is responsible for performing default gateway functionality for its directly attached endpoints. The edge device also takes care of tracking the liveliness of its locally attached discovered endpoints by performing periodic ARP refreshes.

Figure 3-9

Figure 3-9 Anycast Gateway Concept

The scale-out implementation of the distributed anycast gateway provides the default gateway closest to each endpoint. The IP address for the default gateway is shared among all the edge devices, and each edge device is responsible for its respective IP subnet. In addition to the IP address of the default gateway being important, the associated MAC address is equally important. The MAC address is important as each endpoint has a local ARP cache that contains the specific IP-to-MAC binding for the default gateway.

For the host mobility use case, undesirable “black-holing” of traffic may occur if the gateway MAC address changes when a host moves to a server below a different ToR in a different rack even if the default gateway remains the same. To prevent this, the distributed anycast gateways on all the edge devices of the VXLAN EVPN fabric share the same MAC address for the gateway service. This shared MAC address, called the anycast gateway MAC addresses (AGM), is configured to be the same on all the edge devices. In fact, the same AGM is shared across all the different IP subnets, and each subnet has its own unique default gateway IP. The anycast gateway not only provides the most efficient egress routing but also provides direct routing to the VTEP where a given endpoint is attached, thereby eliminating hair-pinning of traffic.

In a situation where an endpoint is silent or undiscovered, no host route information for that endpoint in known to the BGP EVPN control plane. Without the host route information, the next-best route in the routing table should be used so that the packets destined to that endpoint are not dropped. Having each distributed IP anycast gateway advertise a subnet route from each VTEP where that subnet is locally instantiated allows a “next-best” route to be available in the routing table. A remote VTEP, which does not have that subnet instantiated locally, finds that the subnet prefix is reachable over multiple paths via ECMP. When traffic destined to an unknown/silent endpoint in this subnet is received by this remote VTEP, traffic is forwarded to one of the chosen VTEPs that serve the destination subnet based on the calculated ECMP hash. Once the traffic reaches that VTEP, the subnet prefix route is hit, which in turn points to a glean adjacency. This triggers the VTEP to send out an ARP request to the local-connected Layer 2 network, based on the destination subnet information. The ARP request is also sent to the Layer 3 core encapsulated with the Layer 2 VXLAN VNI associated with the destination subnet. As a result, the ARP request reaches the silent endpoint in the specific subnet. The endpoint responds to the ARP request, which in turn gets consumed at the VTEP that is directly attached to that endpoint. This is because all VTEPs share the same anycast gateway MAC. As a result, the endpoint is discovered, and the endpoint information is distributed by this VTEP into the BGP EVPN control plane.

An alternative to the subnet route advertisement approach is to allow the default route ( to achieve similar results with the disadvantage of centralizing the discovery. However, a distributed approach based on subnet prefix routes scales much better to discover the silent endpoints. Figure 3-10 illustrates the logical realization of a distributed IP anycast gateway in a VXLAN BGP EVPN fabric.

Figure 3-10

Figure 3-10 Subnet Prefix Advertisement for Distributed Anycast Gateway

With the distributed IP anycast gateway implementation, any VTEP can service any subnet across the entire VXLAN BGP EVPN fabric. Integrated Routing and Bridging (IRB) from the VXLAN BGP EVPN construct provides the capability to efficiently route and bridge traffic. Regardless of the traffic patterns (east–west or north–south), the hair-pinning of traffic is completely eliminated. The BGP EVPN control plane knows the identity of an endpoint along with the next-hop (VTEP) information that indicates its location. This leads to optimal forwarding in the network without requiring a large Layer 2 domain.

In summary, the AGM across all edge devices and subnets provides seamless host mobility as no changes to the endpoint ARP cache need to occur. This allows for “hot” workload mobility across the entire VXLAN fabric. The distributed anycast gateway moves the Layer 3 gateway to the leafs, thereby providing reduced failure domains, simplified configuration, optimized routing, and transparent endpoint/workload mobility.

Integrated Route and Bridge (IRB)

VXLAN BGP EVPN follows two different semantics for IRB that are documented and published in the IETF’s draft-ietf-bess-evpn-inter-subnet-forwarding (https://tools.ietf.org/html/draft-ietf-bess-evpn-inter-subnet-forwarding).

Asymmetric IRB is the first of the documented first-hop routing operations. It follows a bridge–route–bridge pattern within the local edge device or VTEP. As the name implies, when asymmetric IRB is employed for routing, traffic egressing toward a remote VTEP uses a different VNI than the return traffic from the remote VTEP.

As noted in Figure 3-11, Host A, connected to VTEP V1, wants to communicate to Host X, connected to VTEP V2. Because Host A and Host X belong to different subnets, the asymmetric IRB routing process is used. Post default gateway ARP resolution, Host A sends data traffic toward the default gateway in VLAN 10. From VLAN 10, a routing operation is performed toward VLAN 20, which is mapped to VXLAN VNI 30002. The data traffic is encapsulated into VXLAN with VNI 30002. When the encapsulated traffic arrives on VTEP V2, it is decapsulated and then bridged over toward VLAN 20 because VLAN 20 is also mapped to VNI 30002 on VTEP V2.

Figure 3-11

Figure 3-11 Asymmetric IRB

Host A–to–Host X communication results in a bridge–route–bridge pattern, with the encapsulated traffic traveling with VNI 30002. For the return traffic, Host X sends data traffic to the default gateway for the local subnet corresponding to VLAN 20. After the routing operation from VLAN 20 to VLAN 10 is performed, the traffic is encapsulated with VXLAN VNI 30001 and bridged toward VTEP V1. Once the traffic arrives at VTEP V1, the data traffic is decapsulated and bridged toward VLAN 10 because VLAN 10 is mapped to VNI 30001. Consequently, for return traffic from Host X to Host A, a bridge–route–bridge operation is performed, with encapsulated traffic traveling with VNI 30001. The end-to-end traffic flow from Host A to Host X uses VNI 30002, and the return traffic from Host X to Host A uses VNI 30001.

The preceding example demonstrates traffic asymmetry with different VNIs used for communication between Host A and Host X with asymmetric IRB. Asymmetric IRB requires consistent VNI configuration across all VXLAN VTEP(s) to prevent traffic from being black-holed. Configuration consistency is required for the second bridge operation in the bridge–route–bridge sequence because the sequence fails if there is a missing bridge-domain/VNI configuration corresponding to the network in which the destination resides. Figure 3-12 illustrates the fact that traffic from Host A and Host Y flows fine, but reverse traffic from Host Y to Host A does not work, due to the absence of the configuration for VNI 30001 at the VTEP attached to Host Y. Traffic between Host A and Host X will be correctly forwarded as both IRB interfaces (SVI 10 and SVI 20) are present on the attached VTEP.

Figure 3-12

Figure 3-12 Asymmetric IRB and Inconsistent Provisioning

In addition to asymmetric IRB, another option for IRB operations is symmetric IRB. Whereas asymmetric IRB follows the bridge–route–bridge mode of operation, symmetric IRB follows the bridge–route–route–bridge mode of operation. Symmetric IRB provides additional use cases that are not possible with asymmetric IRB.

Symmetric IRB uses the same forwarding semantics when routing between IP subnets with VRF Lite or MPLS L3VPNs. With symmetric IRB, all traffic egressing and returning from a VTEP uses the same VNI. Specifically, the same Layer 3 VNI (L3VNI) associated with the VRF is used for all routed traffic. The distinctions between the L2VNI and L3VNI are identified in the BGP EVPN control plane specific to the VXLAN header 24-bit VNI field.

As noted in Figure 3-13, Host A is connected to VTEP V1 and wants to communicate with Host Y, attached to VTEP V2. Host A sends data traffic to the default gateway associated with the local subnet in VLAN 10. From VLAN 10, traffic is routed based on the destination IP lookup. The lookup result indicates which traffic needs to be VXLAN encapsulated and sends traffic toward VTEP V2, below which Host Y resides. The encapsulated VXLAN traffic is sent from VTEP V1 to VTEP V2 in VNI 50001, where 50001 is the Layer 3 VNI associated with the VRF in which Host A and Host Y reside. Once the encapsulated VXLAN traffic arrives at VTEP V2, the traffic is decapsulated and routed within the VRF toward VLAN 20 where Host Y resides. In this way, for traffic from Host A to Host Y, a bridge–route–route–bridge symmetric sequence is performed. Note that the Layer 2 VNIs associated with the networks in which Host A and Host Y reside are not used for the routing operation with the symmetric IRB option.

Figure 3-13

Figure 3-13 Symmetric IRB

For the return traffic from Host Y to Host A, the return flow is symmetric, using the same VNI 50001 associated with the VRF. Host Y sends the return traffic to the default gateway associated with its local subnet in VLAN 20. VLAN 20 makes a routing decision toward the VRF and VNI 50001, resulting in the VXLAN-encapsulated traffic being sent to VTEP V1. Once the encapsulated VXLAN traffic arrives at VTEP V1, the traffic is decapsulated and routed toward the VRF and VLAN 10. The return traffic flow from Host Y to Host A follows the same bridge–route–route–bridge sequence, following the same VNI 50001. The traffic from Host A to Host Y and traffic from Host Y to Host A leverage the same VRF VNI 50001, resulting in symmetric traffic flows. In fact, all routed traffic between hosts belonging to different networks within the same VRF employs the same VRF VNI 50001.

Symmetric IRB does not have the requirement of maintaining a consistent configuration across all the edge devices for all the VXLAN networks. In other words, scoped configuration can be implemented when it is not necessary to have all VNIs configured on all the edge devices or VTEPs (see Figure 3-14). However, for a given VRF, the same Layer 3 VNI needs to be configured on all the VTEPs since the VRF enables the sequence of bridge–route–route–bridge operation. For communication between hosts belonging to different VRFs, route leaking is required. For VRF route leaking, an external router or firewall is required to perform VRF-to-VRF communication. It should be noted that support for VRF route leaking requires software support to normalize the control protocol information with the data plane encapsulation. Next to local VRF route leaking, downstream VNI assignment can provide this function for the extranet use cases. These options are discussed further in Chapter 8, “External Connectivity.”

Figure 3-14

Figure 3-14 Symmetric IRB and Inconsistent Provisioning

Both symmetric and asymmetric modes are documented in the IETF draft. The Cisco implementation with NX-OS follows the symmetric IRB mode. The symmetric IRB bridge–route–route–bridge sequence offers flexibility for large-scale multitenant deployments. And the bridge–route–route–bridge sequence follows similar semantics to classic routing across a transit segment. The transit routing segment with VXLAN BGP EVPN is reflected by the Layer 3 VNI associated with the VRF.

Endpoint Mobility

BGP EVPN provides a mechanism to provide endpoint mobility within the fabric. When an endpoint moves, the associated host route prefix is advertised in the control plane through an updated sequence number. The sequence number avoids the need to withdraw and relearn a given prefix during an endpoint move. Instead, an update in the control plane of the new location is performed. Since forwarding in the fabric is always dictated by what is present in the BGP EVPN control plane, traffic is quickly redirected to the updated location of the endpoint that moved, thereby providing a smooth traffic convergence. When an endpoint moves, two host route prefixes for the same endpoint are present in the BGP EVPN control plane. The initial prefix is identified by the original VTEP location, and after the move, the prefix is identified by the new VTEP location.

With two prefixes in the control plane present for the same endpoint, a tiebreaker determines the winner. For this purpose, a BGP extended community called the MAC mobility sequence number is added to the endpoint host route advertisement (specifically Route type 2). The sequence number is updated with every endpoint move. In the event that the sequence number reaches its maximum possible value, the sequence number wraps around and starts over. The MAC mobility sequence number is documented in RFC 7432, which is specific to BGP EVPN.

The endpoint MAC address before the move is learned as a Route type 2 advertisement, and the BGP extended community MAC mobility sequence is set to 0. The value 0 indicates that the MAC address has not had a mobility event, and the endpoint is still at the original location. If a MAC mobility event has been detected, a new Route type 2 (MAC/IP advertisement) is added to the BGP EVPN control plane by the “new” VTEP below which the endpoint moved (its new location). The control plane then sets the MAC mobility sequence number to 1 in the BGP extended community as well as in the new location. There are now two identical MAC/IP advertisements in the BGP EVPN control plane. But only the new advertisement has the MAC mobility sequence number set to a value of 1. All the VTEPs honor this advertisement, thereby sending traffic toward the endpoint at its new location. Figure 3-15 illustrates a sample endpoint mobility use case with the relevant BGP EVPN state where endpoint Host A has moved from VTEP V1 to VTEP V3.

Figure 3-15

Figure 3-15 Endpoint Mobility

An endpoint may move multiple times over the course of its lifetime within a fabric. Every time the endpoint moves, the VTEP that detects its new location increments the sequence number by 1 and then advertises the host prefix for that endpoint into the BGP EVPN control plane. Because the BGP EVPN information is synced across all the VTEPs, every VTEP is aware of whether an endpoint is coming up for the first time or whether it is an endpoint move event based on the previous endpoint reachability information.

An endpoint may be powered off or become unreachable for a variety of reasons. If an endpoint has aged out of the ARP, MAC, and BGP tables, the extended community MAC mobility sequence is 0 as well. If the same endpoint reappears (from the fabric’s and BGP EVPN’s point of view), this reappearance is treated as a new learn event. In other words, the BGP EVPN control plane is aware of the current active endpoints and their respective locations, but the control plane does not store any history of the previous incarnations of the endpoints.

Whether an endpoint actually moves (a hot move) or an endpoint dies and another endpoint assumes its identity (a cold move), from the BGP EVPN control plane’s point of view, the actions taken are similar. This is because the identity of an endpoint in the BGP EVPN control plane is derived from its MAC and/or IP addresses.

An endpoint move event results in either Reverse ARP (RARP) or Gratuitous ARP (GARP), signaling either by the endpoint or on behalf of the endpoint. In the case of RARP, the hypervisor or virtual switch typically sends out a RARP with the SMAC address set to the endpoint MAC address and the DMAC set to the broadcast MAC (FFFF.FFFF.FFFF). In the case of GARP, both the IP and MAC address of the endpoint are notified at the new location. This results in IP/MAC reachability of the endpoint being updated in the BGP EVPN control plane. On reception of this information at the old location (VTEP), an endpoint verification process is performed to determine whether the endpoint has indeed moved away from the old location. Once this validation occurs successfully, the old prefix with the old sequence number is withdrawn from the BGP EVPN control plane. In addition to the prefix being withdrawn from BGP EVPN, the ARP and/or MAC address tables are also cleared of the old location. Figure 3-16 illustrates a sample BGP EVPN output for a prefix associated with an end host that has moved within the fabric. In the figure, the fields of interest, including the MAC Mobility Sequence Number field, are highlighted.

Figure 3-16

Figure 3-16 MAC Mobility Sequence

During the verification procedure, if the endpoint responds at the old location as well, a potential duplicate endpoint is detected since the prefix is being seen at multiple locations. The default detection value for duplicate endpoint detection is “5 moves within 180 seconds.” This means after the fifth move within the time frame of 180 seconds, the VTEP starts a “30-second hold” timer before restarting the verification and cleanup process. After the fifth time (5 moves within 180 seconds), the VTEP freezes the duplicate entry. With the Cisco NX-OS BGP EVPN implementation, these default detection values can be modified via user configuration.

By using the MAC mobility sequence numbers carried with the Route type 2 advertisement (MAC/IP advertisement), the BGP EVPN control plane can identify when a potential location change occurs for an endpoint. When a given prefix is determined to still be reachable, the control plane actively verifies whether a given endpoint has actually moved. The control plane provides a tremendous amount of value in not only verifying and cleaning up mobility events but also detecting duplicates within the VXLAN BGP EVPN network.

Virtual PortChannel (vPC) in VXLAN BGP EVPN

The bundling of multiple physical interfaces into a single logical interface between two chassis is referred to as a port channel, which is also known as a link aggregation group (LAG). Virtual PortChannel (vPC)14 is a technology that provides Layer 2 redundancy across two or more physical chassis. Specifically, a single chassis is connected to two other chassis that are configured as a vPC pair. The industry term for this is Multi-Chassis Link Aggregation Group (MC-LAG). Recall that while the underlying transport network for VXLAN is built with a Layer 3 routing protocol leveraging ECMP, the endpoints still connect to the leafs or edge devices via classic Ethernet. It should be noted that the VXLAN BGP EVPN fabric does not mandate that endpoints must have redundant connections to the fabric. However, in most practical deployments, high availability is typically a requirement, and for that purpose, endpoints are connected to the edge devices via port channels.

Several protocols exist to form port channels, including static-unconditional configurations, protocols such as Port Aggregation Protocol (PAgP),15 and industry standard protocols such as Link Aggregation Control Protocol (LACP).16

Port channels can be configured between a single endpoint and a single network switch. When ToR- or switch-level redundancy (as well as link-level redundancy) is a requirement, an MC-LAG is employed to provide the capability to connect an endpoint to multiple network switches.

vPCs allow interfaces of an endpoint to physically connect to two different network switches. From an endpoint perspective, they see a single network switch connected via a single port channel with multiple links. The endpoint connected to the vPC domain can be a switch, a server, or any other networking device that supports the IEEE 802.3 standard and port channels. vPC allows the creation of Layer 2 port channels that span two switches. vPCs consist of two vPC member switches connected by a peer link, with one being the primary and the other being the secondary. The system formed by the network switches is referred to as a vPC domain (see Figure 3-17).

Figure 3-17

Figure 3-17 Classic Ethernet vPC Domain

The vPC primary and secondary members are initially configured as individual edge devices or VTEPs for VXLAN integration. For northbound traffic, the routed interfaces are part of the underlay network to provide reachability to each individual VTEP. Each VTEP is represented by an individual primary IP address (PIP) per VTEP.

With this in mind, the vPC feature first needs to be enabled, followed by the configuration of the vPC domain between the two vPC member network switches. Next, the vPC peer link (PL) and vPC peer keepalive (PKL) need to be connected between the two nodes. The underlay network also needs to be configured on the vPC peer link for both unicast and multicast (if multicast is used) underlay routing purposes.

A single VTEP is configured to represent the vPC domain in the VXLAN BGP EVPN fabric. To achieve the single VTEP, an anycast VTEP is configured with a common virtual IP address (VIP) that is shared across both switches that form the vPC domain. The anycast IP address is used by all the endpoints behind the vPC domain and is represented by a single anycast VTEP for the vPC domain.

Figure 3-18

Figure 3-18 vPC with VXLAN BGP EVPN

A sample vPC domain is shown in Figure 3-18. VTEP V1 has IP address, and VTEP V2 has IP address These are individual physical IP addresses (PIPs) on the NVE interface or the VTEP. A secondary IP address is added to the two VTEPs, which represents the VIP or anycast IP address (specifically The secondary address depicts the location of any endpoint that is attached to the vPC pair. This address is the next hop advertised in the BGP EVPN control plane advertisement, representing the location of all local endpoints below a given vPC pair of switches.

The VIP is advertised from both vPC member switches so that both vPC members can receive traffic directly from any locally attached endpoints. Remote VTEPs can reach the VIP advertised by both the vPC member switches via ECMP over the underlay routed network. In this way, dual-attached endpoints can be reached as long as there is at least one path to any one of the vPC member switches or VTEPs.

In a vPC domain, both dual-attached endpoints and single-attached endpoints (typically referred to as orphans) are advertised as being reachable via the VIP associated with the anycast VTEP. Consequently, for orphan endpoints below a vPC member switch, a backup path is set up via the vPC peer link if the reachability to the spines is lost due to uplink failures. Hence, an underlay routing adjacency across the vPC peer link is recommended to address the failure scenarios.

If multicast is enabled on the underlay for carrying multidestination traffic, multicast routing (typically PIM) should also be enabled over the peer link. Of note, by default, the advertisement of an anycast VTEP VIP IP address as the next hop with BGP EVPN applies to the MAC/IP advertisements of Route type 2 as well as the IP prefix routes of Route type 5. In this way, all remote VTEPs can always see all BGP EVPN routes behind a vPC pair as being reachable via a single VTEP represented by the VIP. The VTEP associated with the VIP on the VPC peers of a given VPC domain is sometimes referred to as an anycast VTEP for that domain. If there are N leafs deployed in a VXLAN BGP EVPN fabric and all of them are configured as VPC peers, there are N/2 peers present in the network, represented by the respective anycast VTEPs, one per VPC domain.

The term anycast VTEP should not be confused with the anycast gateway. Recall that the anycast gateway refers to the distributed IP anycast gateway for a given subnet that is shared by any and all leafs (including those configured as VPC peers) simultaneously with end host reachability information exchanged between the leafs via BGP EVPN. This is referred to as the gateway because end hosts within a subnet are configured with the default router set to the corresponding anycast gateway IP. Consequently, from an end host point of view, the default gateway’s IP-to-MAC binding should remain the same, regardless of where the end host resides within the fabric. With Cisco NX-OS, all the anycast gateways also share the same globally configured anycast gateway MAC (AGM).

For a given vPC domain, since all ARPs, MACs, and ND entries are synced between the two vPC member switches, both switches are effectively advertising reachability for the same set of prefixes over BGP EVPN. A given vPC member switch thus ignores the BGP EVPN advertisements received from the adjacent vPC member switch since they are part of the same vPC domain because these advertisements are identified by an appropriate site-of-origin extended community attribute that they carry. In this way, only one VTEP or VIP needs to be advertised over BGP EVPN for all endpoints per vPC domain. For a VXLAN BGP EVPN network with N vPC pairs, each remote VTEP needs to know about only N–1 VTEP IPs or VIPs. Even with the presence of only one VIP acting as the anycast VTEP within a given vPC domain, the MP-BGP Route Distinguishers (RDs) are individual identifiers defined on a per-vPC member basis.

However, in some scenarios, IP prefixes may only be advertised by one of the two vPC member switches in a given vPC domain. For example, an individual loopback address may be configured under a VRF on each of the vPC member switches. In this case, because all reachability is advertised into BGP EVPN as being reachable via the anycast VTEP VIP, the traffic destined to a particular vPC member switch may reach its adjacent peer. After decapsulation, the traffic gets black-holed because it will suffer a routing table lookup miss within that VRF.

When IP address information is only advertised by one of the two vPC member switches, such problems may arise. This is because the return traffic needs to reach the correct vPC member switch. Advertising everything from the vPC domain as being reachable via the anycast VTEP IP address does not achieve this. This use case applies to scenarios having southbound IP subnets, orphan-connected IP subnets, individual loopback IP addresses, and/or DHCP relays. With all these use cases, Layer 3 routing adjacency on a per-VRF basis is required to ensure a routing exchange between the vPC domain members. This ensures that even if the packet reaches the incorrect vPC peer, after decapsulation the routing table lookup within the VRF does not suffer a lookup miss.

The configuration of this per-VRF routing adjacency between the vPC member switches may get a bit tedious. However, with BGP EVPN, an elegant solution to achieve similar behavior involves the use of a feature called advertise-pip. This feature is globally enabled on a per-switch basis. Recall that every vPC member switch has a unique PIP and a shared VIP address. With the advertise-pip feature enabled, all IP route prefix reachability is advertised from the individual vPC member switches using the PIP as the next hop in Route type 5 messages. Endpoint reachability corresponding to the IP/MAC addresses via Route type 2 messages continues to use the anycast VTEP or VIP as the next hop (see Figure 3-19). This allows per-switch routing advertisement to be individually advertised, allowing the return traffic to reach the respective VTEP in the vPC domain. In this way, a vPC member switch that is part of a given vPC domain also knows about individual IP prefix reachability of its adjacent vPC peer via BGP EVPN Route type 5 advertisements.

Figure 3-19

Figure 3-19 Advertise PIP with vPC

One additional discussion point pertains to the Router MAC carried in the extended community attributes of the BGP EVPN Route type 2 and Route type 5 messages. With advertise-pip, every vPC switch advertises two VTEPs’ IP addresses (one PIP and one VIP). The PIP uses the switch Router MAC, and the VIP uses a locally derived MAC based on the VIP itself. Both of the vPC member switches derive the same MAC because they are each configured with the same shared VIP under the anycast VTEP. Because the Router MAC extended community is nontransitive, and the VIPs are unique within a given VXLAN BGP EVPN fabric, using locally significant router MACs for the VIPs is not a concern.

In summary, by using the advertise-pip feature with VXLAN BGP EVPN, the next-hop IP address of a VTEP is conditionally handled, depending on whether a Route type 2 (MAC/IP advertisement) or a Route type 5 (IP prefix route) is announced. Advertising PIP allows for efficient handling of vPC-attached endpoints as well as IP subnets specific to the use cases for Layer 2 and Layer 3 border connectivity or DHCP relays. Finally, it should be noted that some differences in the way vPC operates with the different Nexus platforms might exist. For an exhaustive list of the configuration required with VPC in VXLAN BGP EVPN environment for Nexus 9000 platform, please refer to Cisco’s Example of VXLAN BGP EVPN (EBGP).17 And for Nexus 7000 and Nexus 5600 platforms, please refer to Cisco’s Forwarding configurations for Cisco Nexus 5600 and 7000 Series switches in the programmable fabric.18


Dynamic Host Configuration Protocol (DHCP)19 is a commonly used protocol that provides IP address and other options for an endpoint. A DHCP server component is responsible for delivering IP address assignments and managing the binding between an endpoint MAC address and the provided IP address. In addition to distributing the IP address (DHCP scope) itself, additional options such as default gateway, DNS server, and various others (DHCP options) can be assigned to the endpoint that serves as a DHCP client. The requester in the case of DHCP is the DHCP client, which sends a discovery message to the DHCP server. The DHCP server then responds to this discover request with an offer message. Once this initial exchange is complete, a DHCP request (from client to server) is followed by a DHCP acknowledgment, and this completes the assignment of the IP address and subsequent information through the DHCP scope options. Figure 3-20 illustrates this DHCP process.

Figure 3-20

Figure 3-20 DHCP Process

Because the DHCP process relies on broadcasts to exchange information between DHCP client and DHCP server, the DHCP server needs to be in the same Layer 2 network in which the DHCP client resides. This is one deployment method for DHCP. In most cases, multiple IP subnets exist, and routers are in the path between a DHCP client and DHCP server. In this case, a DHCP relay agent is required. The DHCP relay agent is a DHCP protocol helper that snoops DHCP broadcasts and forwards these messages (specifically unicasts) to a specified DHCP server.

The DHCP relay is typically configured on the default gateway facing the DHCP client. The DHCP relay agent can support many different use cases for automated IP addressing, depending on whether the DHCP server resides in the same network, the same VRF, or a different VRF compared to the DHCP client. Having a per-tenant or per-VRF DHCP server is common in certain service provider data centers where strict segregation between tenants is a mandatory requirement. On the other hand, in enterprise environments, having a centralized DHCP server in a shared-services VRF is not uncommon. Traffic crosses tenant (VRF) boundaries with the DHCP server and DHCP client in a different VRF(s). In general, we can differentiate between three DHCP server deployment modes:

  • DHCP client and DHCP server within the same Layer 2 segment (VLAN/L2 VNI)

    • No DHCP relay agent required

    • DHCP discovery from DHCP client is broadcast in the local Layer 2 segment (VLAN/L2 VNI)

    • As DHCP uses broadcasts, multidestination replication is required

  • DHCP client and DHCP server in different IP subnets but in the same tenant (VRF)

    • DHCP relay agent required

    • DHCP discovery from DHCP client is snooped by the relay agent and directed to the DHCP server

  • DHCP client and DHCP server are in different IP subnets and different tenants (VRF)

    • DHCP relay agent required with VRF selection

    • DHCP discovered from DHCP client is snooped by the relay agent and directed to the DHCP server into the VRF where the DHCP server resides

Two of the three DHCP server deployment modes relay data with a DHCP relay agent.20 The DHCP server infrastructure is a shared resource in a multitenant network, and multitenancy support for DHCP services needs to be available. The same DHCP relay agent is responsible for handling the DHCP discovery message and DHCP request message from the DHCP client to the DHCP server and the DHCP offer message and DHCP acknowledgment message from the DHCP server to the requesting DHCP client.

Typically, the DHCP relay agent is configured at the same place where the default gateway IP address is configured for a given subnet. The DHCP relay agent uses the default gateway IP address in the GiAddr field in the DHCP payload for all DHCP messages that are relayed to the DHCP server. The GiAddr field in the relayed DHCP message is used for subnet scope selection so that a free IP address is picked from this subnet at the DHCP server and the respective DCHP options are assigned to the offer that will be returned from the server to the relay agent. The GiAddr field indicates the address to which the DHCP server sends out the response (that is, the DHCP offer or acknowledgment).

Additional considerations need to be taken into account when implementing the DHCP relay agent with the distributed IP anycast gateway in a VXLAN BGP EVPN fabric. This is because the same IP address is shared by all VTEPs that provide Layer 3 service in the form of a default gateway for a given network. Likewise, the same GiAddr field is consequently stamped in DHCP requests relayed by these VTEPs. As a result, the DHCP response from the DHCP server may not reach the same VTEP that had relayed the initial request due to the presence of an anycast IP. In other words, the response may go to any one of the VTEPs servicing the anycast IP, which is clearly undesirable (see Figure 3-21). If a unique IP address per VTEP exists, then that can be used in the GiAddr field, thereby ensuring that the server response is guaranteed to come back to the right VTEP. This is achieved via specification of an ip dhcp relay source-interface xxx command under the Layer 3 interface configured with the anycast gateway IP.

Figure 3-21

Figure 3-21 GiAddr Problem with Distributed IP Anycast Gateway

In this way, the GiAddr field is modified to carry the IP address of the specified source interface. Any given interface on the network switch can be used, as long as the IP address is unique within the network used for reaching the DHCP server. In addition, the IP address must be routable, so that the DHCP server must be able to respond to the individual and unique DHCP relay IP address identified by the GiAddr field. By accomplishing this, the DHCP messages from the client to the server are ensured to travel back and forth through the same network switch or relay agent.

It should be noted that typically the GiAddr field also plays a role in the IP subnet scope selection from which a free IP address is allocated and returned to the DHCP client. Because the GiAddr field has been changed based on the source interface specification, it no longer represents the DHCP scope selection and respective IP address assignment. As a result, a different way to enable scope selection is required. DHCP option 82 (a vendor-specific option) was designed to put in circuit-specific information, and has been used in different deployments (see Figure 3-22). DHCP option 82 has two suboptions that derive the required IP subnet for the IP address assignment, which provides a different approach for DHCP subnet scope selection.

Figure 3-22

Figure 3-22 DHCP Option 82

With DHCP option 82 with the Circuit-ID suboption, the VLAN or VNI information associated with the network in which the client resides is added to the DHCP messages, which provide this information to the DHCP server. The DHCP server needs to support the functionality to select the right DHCP scope, based on the Circuit-ID suboption. DHCP servers that support the DHCP option 82 with the Circuit-ID suboption include Microsoft’s Windows 2012 DHCP server, ISC DHCP server (dhcpd), and Cisco Prime Network Registrar (CPNR).

With the second option, which is the preferred option, the DHCP scope selection is performed using the Link Selection suboption of DHCP option 82. Within the Link Selection suboption, the original client subnet is carried, thereby allowing the correct DHCP scope selection. DHCP servers that support the Link Selection suboption include dhcpd, Infoblox’s DDI, and CPNR. For an exhaustive list of the DHCP relay configuration required in a VXLAN BGP EVPN network, refer to the vPC VTEP DHCP relay configuration example at Cisco.com.21

With the use of the distributed anycast gateway in a VXLAN BGP EVPN fabric and its multitenant capability, centralized DHCP services simplify the ability to provide IP address assignments to endpoints. With a unique source IP address for relaying the DHCP message, and with DHCP option 82, the IP address assignment infrastructure can be centrally provided even across the VRF or tenant boundary.

There are currently no related articles. Please check back later.