An Architecture for Data Center Network Virtualization Overlays (NVO3)

This document presents a high-level architecture for building data center network virtualization overlay (NVO3) networks. The architecture is given at a high-level, showing the major components of an overall system. An important goal is to divide the space into smaller individual components that can be implemented independently and with clear interfaces and interactions with other components. It should be possible to build and implement individual components in isolation and have them work with other components with no changes to other components. That way implementers have flexibility in implementing individual components and can optimize and innovate within their respective components without necessarily requiring changes to other components. The motivation for overlay networks is given in "Problem Statement: Overlays for Network Virtualization" . "Framework for DC Network Virtualization" provides a framework for discussing overlay networks generally and the various components that must work together in building such systems. This document differs from the framework document in that it doesn't attempt to cover all possible approaches within the general design space. Rather, it describes one particular approach that the NVO3 WG has focused on.

This document uses the same terminology as . In addition, the following terms are used: A Network Virtualization Domain is an administrative construct that defines a Network Virtualization Authority (NVA), the set of Network Virtualization Edges (NVEs) associated with that NVA, and the set of virtual networks the NVA manages and supports. NVEs are associated with a (logically centralized) NVA, and an NVE supports communication for any of the virtual networks in the domain. A region over which information about a set of virtual networks is shared. The degenerate case of a single NV Domain corresponds to an NV region corresponding to that domain. The more interesting case occurs when two or more NV Domains share information about part or all of a set of virtual networks that they manage. Two NVAs share information about particular virtual networks for the purpose of supporting connectivity between tenants located in different NV Domains. NVAs can share information about an entire NV domain, or just individual virtual networks. Interface to a Virtual Network as presented to a Tenant System. The TSI logically connects to the NVE via a Virtual Access Point (VAP). To the Tenant System, the TSI is like a Network Interface Card (NIC); the TSI presents itself to a Tenant System as a normal network interface. Unless stated otherwise, the terms VLAN and VLAN Tag are used in this document to denote a C-VLAN and the terms are used interchangeably to improve readability.

Overlay networks are an approach for providing network virtualization services to a set of Tenant Systems (TSs) . With overlays, data traffic between tenants is tunneled across the underlying data center's IP network. The use of tunnels provides a number of benefits by decoupling the network as viewed by tenants from the underlying physical network across which they communicate. Tenant Systems connect to Virtual Networks (VNs), with each VN having associated attributes defining properties of the network, such as the set of members that connect to it. Tenant Systems connected to a virtual network typically communicate freely with other Tenant Systems on the same VN, but communication between Tenant Systems on one VN and those external to the VN (whether on another VN or connected to the Internet) is carefully controlled and governed by policy. The NVO3 architecture does not impose any restrictions to the application of policy controls even within a VN. A Network Virtualization Edge (NVE) is the entity that implements the overlay functionality. An NVE resides at the boundary between a Tenant System and the overlay network as shown in . An NVE creates and maintains local state about each Virtual Network for which it is providing service on behalf of a Tenant System.

The following subsections describe key aspects of an overlay system in more detail. describes the service model (Ethernet vs. IP) provided to Tenant Systems. describes NVEs in more detail. introduces the Network Virtualization Authority, from which NVEs obtain information about virtual networks. provides background on Virtual Machine (VM) orchestration systems and their use of virtual networks.

A Virtual Network provides either L2 or L3 service to connected tenants. For L2 service, VNs transport Ethernet frames, and a Tenant System is provided with a service that is analogous to being connected to a specific L2 C-VLAN. L2 broadcast frames are generally delivered to all (and multicast frames delivered to a subset of) the other Tenant Systems on the VN. To a Tenant System, it appears as if they are connected to a regular L2 Ethernet link. Within the NVO3 architecture, tenant frames are tunneled to remote NVEs based on the MAC addresses of the frame headers as originated by the Tenant System. On the underlay, NVO3 packets are forwarded between NVEs based on the outer addresses of tunneled packets. For L3 service, VNs transport IP datagrams, and a Tenant System is provided with a service that only supports IP traffic. Within the NVO3 architecture, tenant frames are tunneled to remote NVEs based on the IP addresses of the packet originated by the Tenant System; any L2 destination addresses provided by Tenant Systems are effectively ignored by the NVEs and overlay network. For L3 service, the Tenant System will be configured with an IP subnet that is effectively a point-to-point link, i.e., having only the Tenant System and a next-hop router address on it. L2 service is intended for systems that need native L2 Ethernet service and the ability to run protocols directly over Ethernet (i.e., not based on IP). L3 service is intended for systems in which all the traffic can safely be assumed to be IP. It is important to note that whether an NVO3 network provides L2 or L3 service to a Tenant System, the Tenant System does not generally need to be aware of the distinction. In both cases, the virtual network presents itself to the Tenant System as an L2 Ethernet interface. An Ethernet interface is used in both cases simply as a widely supported interface type that essentially all Tenant Systems already support. Consequently, no special software is needed on Tenant Systems to use an L3 vs. an L2 overlay service. NVO3 can also provide a combined L2 and L3 service to tenants. A combined service provides L2 service for intra-VN communication, but also provides L3 service for L3 traffic entering or leaving the VN. Architecturally, the handling of a combined L2/L3 service within the NVO3 architecture is intended to match what is commonly done today in non-overlay environments by devices providing a combined bridge/router service. With combined service, the virtual network itself retains the semantics of L2 service and all traffic is processed according to its L2 semantics. In addition, however, traffic requiring IP processing is also processed at the IP level. The IP processing for a combined service can be implemented on a standalone device attached to the virtual network (e.g., an IP router) or implemented locally on the NVE (see Section 5.4 on Distributed Gateways). For unicast traffic, NVE implementation of a combined service may result in a packet being delivered to another TS attached to the same NVE (on either the same or a different VN) or tunneled to a remote NVE, or even forwarded outside the NV domain. For multicast or broadcast packets, the combination of NVE L2 and L3 processing may result in copies of the packet receiving both L2 and L3 treatments to realize delivery to all of the destinations involved. This distributed NVE implementation of IP routing results in the same network delivery behavior as if the L2 processing of the packet included delivery of the packet to an IP router attached to the L2 VN as a TS, with the router having additional network attachments to other networks, either virtual or not.

An NVO3 L2 virtual network service may include encapsulated L2 VLAN tags provided by a Tenant System, but does not use encapsulated tags in deciding where and how to forward traffic. Such VLAN tags can be passed through, so that Tenant Systems that send or expect to receive them can be supported as appropriate. The processing of VLAN tags that an NVE receives from a TS is controlled by settings associated with the VAP. Just as in the case with ports on Ethernet switches, a number of settings are possible. For example, C-TAGs can be passed through transparently, they could always be stripped upon receipt from a Tenant System, they could be compared against a list of explicitly configured tags, etc. Note that that there are additional considerations when VLAN tags are used to identify both the VN and a Tenant System VLAN within that VN, as described in below.

For L3 service, Tenant Systems should expect the IPv4 TTL (Time to Live) or IPv6 Hop Limit in the packets they send to be decremented by at least 1. For L2 service, neither the TTL nor the Hop Limit (when the packet is IP) are modified. The underlay network manages TTLs and Hop Limits in the outer IP encapsulation - the values in these fields could be independent from or related to the values in the same fields of tenant IP packets.

Tenant Systems connect to NVEs via a Tenant System Interface (TSI). The TSI logically connects to the NVE via a Virtual Access Point (VAP) and each VAP is associated with one Virtual Network as shown in . To the Tenant System, the TSI is like a NIC; the TSI presents itself to a Tenant System as a normal network interface. On the NVE side, a VAP is a logical network port (virtual or physical) into a specific virtual network. Note that two different Tenant Systems (and TSIs) attached to a common NVE can share a VAP (e.g., TS1 and TS2 in ) so long as they connect to the same Virtual Network.

The Overlay Module performs the actual encapsulation and decapsulation of tunneled packets. The NVE maintains state about the virtual networks it is a part of so that it can provide the Overlay Module with such information as the destination address of the NVE to tunnel a packet to and the Context ID that should be placed in the encapsulation header to identify the virtual network that a tunneled packet belongs to. On the data center network side, the NVE sends and receives native IP traffic. When ingressing traffic from a Tenant System, the NVE identifies the egress NVE to which the packet should be sent, adds an overlay encapsulation header, and sends the packet on the underlay network. When receiving traffic from a remote NVE, an NVE strips off the encapsulation header, and delivers the (original) packet to the appropriate Tenant System. When the source and destination Tenant System are on the same NVE, no encapsulation is needed and the NVE forwards traffic directly. Conceptually, the NVE is a single entity implementing the NVO3 functionality. In practice, there are a number of different implementation scenarios, as described in detail in .

Address dissemination refers to the process of learning, building and distributing the mapping/forwarding information that NVEs need in order to tunnel traffic to each other on behalf of communicating Tenant Systems. For example, in order to send traffic to a remote Tenant System, the sending NVE must know the destination NVE for that Tenant System. One way to build and maintain mapping tables is to use learning, as 802.1 bridges do . When forwarding traffic to multicast or unknown unicast destinations, an NVE could simply flood traffic. While flooding works, it can lead to traffic hot spots and can lead to problems in larger networks (e.g., excessive amounts of flooded traffic). Alternatively, to reduce the scope of where flooding must take place, or to eliminate it all together, NVEs can make use of a Network Virtualization Authority (NVA). An NVA is the entity that provides address mapping and other information to NVEs. NVEs interact with an NVA to obtain any required address mapping information they need in order to properly forward traffic on behalf of tenants. The term NVA refers to the overall system, without regards to its scope or how it is implemented. NVAs provide a service, and NVEs access that service via an NVE-to-NVA protocol as discussed in . Even when an NVA is present, Ethernet bridge MAC address learning could be used as a fallback mechanism, should the NVA be unable to provide an answer or for other reasons. This document does not consider flooding approaches in detail, as there are a number of benefits in using an approach that depends on the presence of an NVA. For the rest of this document, it is assumed that an NVA exists and will be used. NVAs are discussed in more detail in .

VM orchestration systems manage server virtualization across a set of servers. Although VM management is a separate topic from network virtualization, the two areas are closely related. Managing the creation, placement, and movement of VMs also involves creating, attaching to and detaching from virtual networks. A number of existing VM orchestration systems have incorporated aspects of virtual network management into their systems. Note also, that although this section uses the term "VM" and "hypervisor" throughout, the same issues apply to other virtualization approaches, including Linux Containers (LXC), BSD Jails, Network Service Appliances as discussed in Section 5.1, etc.. From an NVO3 perspective, it should be assumed that where the document uses the term "VM" and "hypervisor", the intention is that the discussion also applies to other systems, where, e.g., the host operating system plays the role of the hypervisor in supporting virtualization, and a container plays the equivalent role as a VM. When a new VM image is started, the VM orchestration system determines where the VM should be placed, interacts with the hypervisor on the target server to load and start the VM and controls when a VM should be shutdown or migrated elsewhere. VM orchestration systems also have knowledge about how a VM should connect to a network, possibly including the name of the virtual network to which a VM is to connect. The VM orchestration system can pass such information to the hypervisor when a VM is instantiated. VM orchestration systems have significant (and sometimes global) knowledge over the domain they manage. They typically know on what servers a VM is running, and meta data associated with VM images can be useful from a network virtualization perspective. For example, the meta data may include the addresses (MAC and IP) the VMs will use and the name(s) of the virtual network(s) they connect to. VM orchestration systems run a protocol with an agent running on the hypervisor of the servers they manage. That protocol can also carry information about what virtual network a VM is associated with. When the orchestrator instantiates a VM on a hypervisor, the hypervisor interacts with the NVE in order to attach the VM to the virtual networks it has access to. In general, the hypervisor will need to communicate significant VM state changes to the NVE. In the reverse direction, the NVE may need to communicate network connectivity information back to the hypervisor. Examples of deployed VM orchestration systems include VMware's vCenter Server, Microsoft's System Center Virtual Machine Manager, and systems based on OpenStack and its associated plugins (e.g., Nova and Neutron). Each can pass information about what virtual networks a VM connects to down to the hypervisor. The protocol used between the VM orchestration system and hypervisors is generally proprietary. It should be noted that VM orchestration systems may not have direct access to all networking related information a VM uses. For example, a VM may make use of additional IP or MAC addresses that the VM management system is not aware of.

As introduced in an NVE is the entity that implements the overlay functionality. This section describes NVEs in more detail. An NVE will have two external interfaces: On the Tenant System facing side, an NVE interacts with the hypervisor (or equivalent entity) to provide the NVO3 service. An NVE will need to be notified when a Tenant System "attaches" to a virtual network (so it can validate the request and set up any state needed to send and receive traffic on behalf of the Tenant System on that VN). Likewise, an NVE will need to be informed when the Tenant System "detaches" from the virtual network so that it can reclaim state and resources appropriately. On the data center network facing side, an NVE interfaces with the data center underlay network, sending and receiving tunneled packets to and from the underlay. The NVE may also run a control protocol with other entities on the network, such as the Network Virtualization Authority.

When server virtualization is used, the entire NVE functionality will typically be implemented as part of the hypervisor and/or virtual switch on the server. In such cases, the Tenant System interacts with the hypervisor and the hypervisor interacts with the NVE. Because the interaction between the hypervisor and NVE is implemented entirely in software on the server, there is no "on-the-wire" protocol between Tenant Systems (or the hypervisor) and the NVE that needs to be standardized. While there may be APIs between the NVE and hypervisor to support necessary interaction, the details of such an API are not in-scope for the NVO3 WG at the time of publication of this memo. Implementing NVE functionality entirely on a server has the disadvantage that server CPU resources must be spent implementing the NVO3 functionality. Experimentation with overlay approaches and previous experience with TCP and checksum adapter offloads suggests that offloading certain NVE operations (e.g., encapsulation and decapsulation operations) onto the physical network adapter can produce performance advantages. As has been done with checksum and/or TCP server offload and other optimization approaches, there may be benefits to offloading common operations onto adapters where possible. Just as important, the addition of an overlay header can disable existing adapter offload capabilities that are generally not prepared to handle the addition of a new header or other operations associated with an NVE. While the exact details of how to split the implementation of specific NVE functionality between a server and its network adapters is an implementation matter and outside the scope of IETF standardization, the NVO3 architecture should be cognizant of and support such separation. Ideally, it may even be possible to bypass the hypervisor completely on critical data path operations so that packets between a TS and its VN can be sent and received without having the hypervisor involved in each individual packet operation.

Another possible scenario leads to the need for a split NVE implementation. An NVE running on a server (e.g. within a hypervisor) could support NVO3 service towards the tenant, but not perform all NVE functions (e.g., encapsulation) directly on the server; some of the actual NVO3 functionality could be implemented on (i.e., offloaded to) an adjacent switch to which the server is attached. While one could imagine a number of link types between a server and the NVE, one simple deployment scenario would involve a server and NVE separated by a simple L2 Ethernet link. A more complicated scenario would have the server and NVE separated by a bridged access network, such as when the NVE resides on a ToR, with an embedded switch residing between servers and the ToR. For the split NVE case, protocols will be needed that allow the hypervisor and NVE to negotiate and setup the necessary state so that traffic sent across the access link between a server and the NVE can be associated with the correct virtual network instance. Specifically, on the access link, traffic belonging to a specific Tenant System would be tagged with a specific VLAN C-TAG that identifies which specific NVO3 virtual network instance it connects to. The hypervisor-NVE protocol would negotiate which VLAN C-TAG to use for a particular virtual network instance. More details of the protocol requirements for functionality between hypervisors and NVEs can be found in .

Preserving tenant VLAN tags across an NVO3 VN as described in poses additional complications in the split-NVE case. The portion of the NVE that performs the encapsulation function needs access to the specific VLAN tags that the Tenant System is using in order to include them in the encapsulated packet. When an NVE is implemented entirely within the hypervisor, the NVE has access to the complete original packet (including any VLAN tags) sent by the tenant. In the split-NVE case, however, the VLAN tag used between the hypervisor and offloaded portions of the NVE normally only identify the specific VN that traffic belongs to. In order to allow a tenant to preserve VLAN information from end to end between TS when in the split-NVE case, additional mechanisms would be needed (e.g., carry an additional VLAN tag by carrying both a C-Tag and an S-Tag as specified in where the C-Tag identifies the TS VLAN end-to-end and the S-Tag identifies the VN locally between each TS and the corresponding NVE).

NVEs maintain internal data structures and state to support the sending and receiving of tenant traffic. An NVE may need some or all of the following information: An NVE keeps track of which attached Tenant Systems are connected to which virtual networks. When a Tenant System attaches to a virtual network, the NVE will need to create or update local state for that virtual network. When the last Tenant System detaches from a given VN, the NVE can reclaim state associated with that VN. For tenant unicast traffic, an NVE maintains a per-VN table of mappings from Tenant System (inner) addresses to remote NVE (outer) addresses. For tenant multicast (or broadcast) traffic, an NVE maintains a per-VN table of mappings and other information on how to deliver tenant multicast (or broadcast) traffic. If the underlying network supports IP multicast, the NVE could use IP multicast to deliver tenant traffic. In such a case, the NVE would need to know what IP underlay multicast address to use for a given VN. Alternatively, if the underlying network does not support multicast, a source NVE could use unicast replication to deliver traffic. In such a case, an NVE would need to know which remote NVEs are participating in the VN. An NVE could use both approaches, switching from one mode to the other depending on such factors as bandwidth efficiency and group membership sparseness. discusses the subject of multicast handling in NVO3 in further detail. An NVE maintains necessary information to encapsulate outgoing traffic, including what type of encapsulation and what value to use for a Context ID to identify the VN within the encapsulation header. In order to deliver incoming encapsulated packets to the correct Tenant Systems, an NVE maintains the necessary information to map incoming traffic to the appropriate VAP (i.e., Tenant System Interface). An NVE may find it convenient to maintain additional per-VN information such as QoS settings, Path MTU information, ACLs, etc.

NVEs may be multi-homed. That is, an NVE may have more than one IP address associated with it on the underlay network. Multihoming happens in two different scenarios. First, an NVE may have multiple interfaces connecting it to the underlay. Each of those interfaces will typically have a different IP address, resulting in a specific Tenant Address (on a specific VN) being reachable through the same NVE but through more than one underlay IP address. Second, a specific tenant system may be reachable through more than one NVE, each having one or more underlay addresses. In both cases, NVE address mapping functionality needs to support one-to-many mappings and enable a sending NVE to (at a minimum) be able to fail over from one IP address to another, e.g., should a specific NVE underlay address become unreachable. Finally, multi-homed NVEs introduce complexities when source unicast replication is used to implement tenant multicast as described in Section 4.3. Specifically, an NVE should only receive one copy of a replicated packet. Multi-homing is needed to support important use cases. First, a bare metal server may have multiple uplink connections to either the same or different NVEs. Having only a single physical path to an upstream NVE, or indeed, having all traffic flow through a single NVE would be considered unacceptable in highly-resilient deployment scenarios that seek to avoid single points of failure. Moreover, in today's networks, the availability of multiple paths would require that they be usable in an active-active fashion (e.g., for load balancing).

The VAP is the NVE-side of the interface between the NVE and the TS. Traffic to and from the tenant flows through the VAP. If an NVE runs into difficulties sending traffic received on the VAP, it may need to signal such errors back to the VAP. Because the VAP is an emulation of a physical port, its ability to signal NVE errors is limited and lacks sufficient granularity to reflect all possible errors an NVE may encounter (e.g., inability reach a particular destination). Some errors, such as an NVE losing all of its connections to the underlay, could be reflected back to the VAP by effectively disabling it. This state change would reflect itself on the TS as an interface going down, allowing the TS to implement interface error handling, e.g., failover, in the same manner as when a physical interfaces becomes disabled.

This section describes a number of special Tenant System types and how they fit into an NVO3 system.

Some Network Service Appliances (virtual or physical) provide tenant-aware services. That is, the specific service they provide depends on the identity of the tenant making use of the service. For example, firewalls are now becoming available that support multi-tenancy where a single firewall provides virtual firewall service on a per-tenant basis, using per-tenant configuration rules and maintaining per-tenant state. Such appliances will be aware of the VN an activity corresponds to while processing requests. Unlike server virtualization, which shields VMs from needing to know about multi-tenancy, a Network Service Appliance may explicitly support multi-tenancy. In such cases, the Network Service Appliance itself will be aware of network virtualization and either embed an NVE directly, or implement a split NVE as described in . Unlike server virtualization, however, the Network Service Appliance may not be running a hypervisor and the VM orchestration system may not interact with the Network Service Appliance. The NVE on such appliances will need to support a control plane to obtain the necessary information needed to fully participate in an NV Domain.

Many data centers will continue to have at least some servers operating as non-virtualized (or "bare metal") machines running a traditional operating system and workload. In such systems, there will be no NVE functionality on the server, and the server will have no knowledge of NVO3 (including whether overlays are even in use). In such environments, the NVE functionality can reside on the first-hop physical switch. In such a case, the network administrator would (manually) configure the switch to enable the appropriate NVO3 functionality on the switch port connecting the server and associate that port with a specific virtual network. Such configuration would typically be static, since the server is not virtualized, and once configured, is unlikely to change frequently. Consequently, this scenario does not require any protocol or standards work.

Gateways on VNs relay traffic onto and off of a virtual network. Tenant Systems use gateways to reach destinations outside of the local VN. Gateways receive encapsulated traffic from one VN, remove the encapsulation header, and send the native packet out onto the data center network for delivery. Outside traffic enters a VN in a reverse manner. Gateways can be either virtual (i.e., implemented as a VM) or physical (i.e., as a standalone physical device). For performance reasons, standalone hardware gateways may be desirable in some cases. Such gateways could consist of a simple switch forwarding traffic from a VN onto the local data center network, or could embed router functionality. On such gateways, network interfaces connecting to virtual networks will (at least conceptually) embed NVE (or split-NVE) functionality within them. As in the case with Network Service Appliances, gateways may not support a hypervisor and will need an appropriate control plane protocol to obtain the information needed to provide NVO3 service. Gateways handle several different use cases. For example, one use case consists of systems supporting overlays together with systems that do not (e.g., bare metal servers). Gateways could be used to connect legacy systems supporting, e.g., L2 VLANs, to specific virtual networks, effectively making them part of the same virtual network. Gateways could also forward traffic between a virtual network and other hosts on the data center network or relay traffic between different VNs. Finally, gateways can provide external connectivity such as Internet or VPN access.

As can be seen from the discussion above, there are several types of gateways that can exist in an NVO3 environment. This section breaks them down into the various types that could be supported. Note that each of the types below could be implemented in either a centralized manner or distributed to co-exist with the NVEs.

L2 Gateways act as layer 2 bridges to forward Ethernet frames based on the MAC addresses present in them. This type of gateway bridges traffic between L2 VNs and other legacy L2 networks such as VLANs or L2 VPNs. The main motivation for this type of gateway to create separate groups of Tenant Systems using L2 VNs such that the gateway can enforce network policies between each L2 VN.

L3 Gateways forward IP packets based on the IP addresses present in the packets. This type of gateway forwards packets on between L3 VNs and legacy L2 networks such as VLANs or L2 VPNs. The MAC address in any frames forwarded between the legacy L2 network would be the MAC address of the gateway. The type of gateway forwards packets between L3 VNs and legacy L3 networks. These legacy L3 networks could be local the data center, in the WAN, or an L3 VPN. This type of gateway forwards packets on between L3 VNs and L2 VNs. The MAC address in any frames forwarded between the L2 VN would be the MAC address of the gateway. This type of gateway acts similar to a traditional router that forwards between L2 interfaces. The MAC address in any frames forwarded between the L2 VNs would be the MAC address of the gateway. The main motivation for this type of gateway to create separate groups of Tenant Systems using L3 VNs such that the gateway can enforce network policies between each L3 VN.

The relaying of traffic from one VN to another deserves special consideration. Whether traffic is permitted to flow from one VN to another is a matter of policy, and would not (by default) be allowed unless explicitly enabled. In addition, NVAs are the logical place to maintain policy information about allowed inter-VN communication. Policy enforcement for inter-VN communication can be handled in (at least) two different ways. Explicit gateways could be the central point for such enforcement, with all inter-VN traffic forwarded to such gateways for processing. Alternatively, the NVA can provide such information directly to NVEs, by either providing a mapping for a target TS on another VN, or indicating that such communication is disallowed by policy. When inter-VN gateways are centralized, traffic between TSs on different VNs can take suboptimal paths, i.e., triangular routing results in paths that always traverse the gateway. In the worst case, traffic between two TSs connected to the same NVE can be hair-pinned through an external gateway. As an optimization, individual NVEs can be part of a distributed gateway that performs such relaying, reducing or completely eliminating triangular routing. In a distributed gateway, each ingress NVE can perform such relaying activity directly, so long as it has access to the policy information needed to determine whether cross-VN communication is allowed. Having individual NVEs be part of a distributed gateway allows them to tunnel traffic directly to the destination NVE without the need to take suboptimal paths. The NVO3 architecture supports distributed gateways for the case of inter-VN communication. Such support requires that NVO3 control protocols include mechanisms for the maintenance and distribution of policy information about what type of cross-VN communication is allowed so that NVEs acting as distributed gateways can tunnel traffic from one VN to another as appropriate. Distributed gateways could also be used to distribute other traditional router services to individual NVEs. The NVO3 architecture does not preclude such implementations, but does not define or require them as they are outside the scope of the NVO3 architecture.

For an L2 service, strictly speaking, special processing of Address Resolution Protocol (ARP) (and IPv6 Neighbor Discovery (ND) ) is not required. ARP requests are broadcast, and an NVO3 can deliver ARP requests to all members of a given L2 virtual network, just as it does for any packet sent to an L2 broadcast address. Similarly, ND requests are sent via IP multicast, which NVO3 can support by delivering via L2 multicast. However, as a performance optimization, an NVE can intercept ARP (or ND) requests from its attached TSs and respond to them directly using information in its mapping tables. Since an NVE will have mechanisms for determining the NVE address associated with a given TS, the NVE can leverage the same mechanisms to suppress sending ARP and ND requests for a given TS to other members of the VN. The NVO3 architecture supports such a capability.

Individual NVEs will interact with each other for the purposes of tunneling and delivering traffic to remote TSs. At a minimum, a control protocol may be needed for tunnel setup and maintenance. For example, tunneled traffic may need to be encrypted or integrity protected, in which case it will be necessary to set up appropriate security associations between NVE peers. It may also be desirable to perform tunnel maintenance (e.g., continuity checks) on a tunnel in order to detect when a remote NVE becomes unreachable. Such generic tunnel setup and maintenance functions are not generally NVO3-specific. Hence, the NVO3 architecture expects to leverage existing tunnel maintenance protocols rather than defining new ones. Some NVE-NVE interactions may be specific to NVO3 (and in particular be related to information kept in mapping tables) and agnostic to the specific tunnel type being used. For example, when tunneling traffic for TS-X to a remote NVE, it is possible that TS-X is not presently associated with the remote NVE. Normally, this should not happen, but there could be race conditions where the information an NVE has learned from the NVA is out-of-date relative to actual conditions. In such cases, the remote NVE could return an error or warning indication, allowing the sending NVE to attempt a recovery or otherwise attempt to mitigate the situation. The NVE-NVE interaction could signal a range of indications, for example: "No such TS here", upon a receipt of a tunneled packet for an unknown TS. "TS-X not here, try the following NVE instead" (i.e., a redirect). Delivered to correct NVE, but could not deliver packet to TS-X (soft error). Delivered to correct NVE, but could not deliver packet to TS-X (hard error). When an NVE receives information from a remote NVE that conflicts with the information it has in its own mapping tables, it should consult with the NVA to resolve those conflicts. In particular, it should confirm that the information it has is up-to-date, and it might indicate the error to the NVA, so as to nudge the NVA into following up (as appropriate). While it might make sense for an NVE to update its mapping table temporarily in response to an error from a remote NVE, any changes must be handled carefully as doing so can raise security considerations if the received information cannot be authenticated. That said, a sending NVE might still take steps to mitigate a problem, such as applying rate limiting to data traffic towards a particular NVE or TS.

Before sending to and receiving traffic from a virtual network, an NVE must obtain the information needed to build its internal forwarding tables and state as listed in . An NVE can obtain such information from a Network Virtualization Authority. The Network Virtualization Authority (NVA) is the entity that is expected to provide address mapping and other information to NVEs. NVEs can interact with an NVA to obtain any required information they need in order to properly forward traffic on behalf of tenants. The term NVA refers to the overall system, without regards to its scope or how it is implemented.

There are two primary ways in which an NVA can obtain the address dissemination information it manages. The NVA can obtain information either from the VM orchestration system, and/or directly from the NVEs themselves. On virtualized systems, the NVA may be able to obtain the address mapping information associated with VMs from the VM orchestration system itself. If the VM orchestration system contains a master database for all the virtualization information, having the NVA obtain information directly to the orchestration system would be a natural approach. Indeed, the NVA could effectively be co-located with the VM orchestration system itself. In such systems, the VM orchestration system communicates with the NVE indirectly through the hypervisor. However, as described in not all NVEs are associated with hypervisors. In such cases, NVAs cannot leverage VM orchestration protocols to interact with an NVE and will instead need to peer directly with them. By peering directly with an NVE, NVAs can obtain information about the TSs connected to that NVE and can distribute information to the NVE about the VNs those TSs are associated with. For example, whenever a Tenant System attaches to an NVE, that NVE would notify the NVA that the TS is now associated with that NVE. Likewise when a TS detaches from an NVE, that NVE would inform the NVA. By communicating directly with NVEs, both the NVA and the NVE are able to maintain up-to-date information about all active tenants and the NVEs to which they are attached.

For reliability and fault tolerance reasons, an NVA would be implemented in a distributed or replicated manner without single points of failure. How the NVA is implemented, however, is not important to an NVE so long as the NVA provides a consistent and well-defined interface to the NVE. For example, an NVA could be implemented via database techniques whereby a server stores address mapping information in a traditional (possibly replicated) database. Alternatively, an NVA could be implemented in a distributed fashion using an existing (or modified) routing protocol to maintain and distribute mappings. So long as there is a clear interface between the NVE and NVA, how an NVA is architected and implemented is not important to an NVE. A number of architectural approaches could be used to implement NVAs themselves. NVAs manage address bindings and distribute them to where they need to go. One approach would be to use Border Gateway Protocol (BGP) (possibly with extensions) and route reflectors. Another approach could use a transaction-based database model with replicated servers. Because the implementation details are local to an NVA, there is no need to pick exactly one solution technology, so long as the external interfaces to the NVEs (and remote NVAs) are sufficiently well defined to achieve interoperability.

Conceptually, from the perspective of an NVE, an NVA is a single entity. An NVE interacts with the NVA, and it is the NVA's responsibility for ensuring that interactions between the NVE and NVA result in consistent behavior across the NVA and all other NVEs using the same NVA. Because an NVA is built from multiple internal components, an NVA will have to ensure that information flows to all internal NVA components appropriately. One architectural question is how the NVA presents itself to the NVE. For example, an NVA could be required to provide access via a single IP address. If NVEs only have one IP address to interact with, it would be the responsibility of the NVA to handle NVA component failures, e.g., by using a "floating IP address" that migrates among NVA components to ensure that the NVA can always be reached via the one address. Having all NVA accesses through a single IP address, however, adds constraints to implementing robust failover, load balancing, etc. In the NVO3 architecture, an NVA is accessed through one or more IP addresses (or IP address/port combination). If multiple IP addresses are used, each IP address provides equivalent functionality, meaning that an NVE can use any of the provided addresses to interact with the NVA. Should one address stop working, an NVE is expected to failover to another. While the different addresses result in equivalent functionality, one address may respond more quickly than another, e.g., due to network conditions, load on the server, etc. To provide some control over load balancing, NVA addresses may have an associated priority. Addresses are used in order of priority, with no explicit preference among NVA addresses having the same priority. To provide basic load-balancing among NVAs of equal priorities, NVEs could use some randomization input to select among equal-priority NVAs. Such a priority scheme facilitates failover and load balancing, for example, allowing a network operator to specify a set of primary and backup NVAs. It may be desirable to have individual NVA addresses responsible for a subset of information about an NV Domain. In such a case, NVEs would use different NVA addresses for obtaining or updating information about particular VNs or TS bindings. A key question with such an approach is how information would be partitioned, and how an NVE could determine which address to use to get the information it needs. Another possibility is to treat the information on which NVA addresses to use as cached (soft-state) information at the NVEs, so that any NVA address can be used to obtain any information, but NVEs are informed of preferences for which addresses to use for particular information on VNs or TS bindings. That preference information would be cached for future use to improve behavior - e.g., if all requests for a specific subset of VNs are forwarded to a specific NVA component, the NVE can optimize future requests within that subset by sending them directly to that NVA component via its address.

As outlined in , an NVE needs certain information in order to perform its functions. To obtain such information from an NVA, an NVE-to-NVA protocol is needed. The NVE-to-NVA protocol provides two functions. First it allows an NVE to obtain information about the location and status of other TSs with which it needs to communicate. Second, the NVE-to-NVA protocol provides a way for NVEs to provide updates to the NVA about the TSs attached to that NVE (e.g., when a TS attaches or detaches from the NVE), or about communication errors encountered when sending traffic to remote NVEs. For example, an NVE could indicate that a destination it is trying to reach at a destination NVE is unreachable for some reason. While having a direct NVE-to-NVA protocol might seem straightforward, the existence of existing VM orchestration systems complicates the choices an NVE has for interacting with the NVA.

An NVE interacts with an NVA in at least two (quite different) ways: NVEs embedded within the same server as the hypervisor can obtain necessary information entirely through the hypervisor-facing side of the NVE. Such an approach is a natural extension to existing VM orchestration systems supporting server virtualization because an existing protocol between the hypervisor and VM orchestration system already exists and can be leveraged to obtain any needed information. Specifically, VM orchestration systems used to create, terminate and migrate VMs already use well-defined (though typically proprietary) protocols to handle the interactions between the hypervisor and VM orchestration system. For such systems, it is a natural extension to leverage the existing orchestration protocol as a sort of proxy protocol for handling the interactions between an NVE and the NVA. Indeed, existing implementations can already do this. Alternatively, an NVE can obtain needed information by interacting directly with an NVA via a protocol operating over the data center underlay network. Such an approach is needed to support NVEs that are not associated with systems performing server virtualization (e.g., as in the case of a standalone gateway) or where the NVE needs to communicate directly with the NVA for other reasons. The NVO3 architecture will focus on support for the second model above. Existing virtualization environments are already using the first model. But they are not sufficient to cover the case of standalone gateways -- such gateways may not support virtualization and do not interface with existing VM orchestration systems.

An NVE can interact directly with an NVA via an NVE-to-NVA protocol. Such a protocol can be either independent of the NVA internal protocol, or an extension of it. Using a purpose-specific protocol would provide architectural separation and independence between the NVE and NVA. The NVE and NVA interact in a well-defined way, and changes in the NVA (or NVE) do not need to impact each other. Using a dedicated protocol also ensures that both NVE and NVA implementations can evolve independently and without dependencies on each other. Such independence is important because the upgrade path for NVEs and NVAs is quite different. Upgrading all the NVEs at a site will likely be more difficult in practice than upgrading NVAs because of their large number - one on each end device. In practice, it would be prudent to assume that once an NVE has been implemented and deployed, it may be challenging to get subsequent NVE extensions and changes implemented and deployed, whereas an NVA (and its associated internal protocols) are more likely to evolve over time as experience is gained from usage and upgrades will involve fewer nodes. Requirements for a direct NVE-NVA protocol can be found in

Information flows between NVEs and NVAs in both directions. The NVA maintains information about all VNs in the NV Domain, so that NVEs do not need to do so themselves. NVEs obtain from the NVA information about where a given remote TS destination resides. NVAs in turn obtain information from NVEs about the individual TSs attached to those NVEs. While the NVA could push information relevant to every virtual network to every NVE, such an approach scales poorly and is unnecessary. In practice, a given NVE will only need and want to know about VNs to which it is attached. Thus, an NVE should be able to subscribe to updates only for the virtual networks it is interested in receiving updates for. The NVO3 architecture supports a model where an NVE is not required to have full mapping tables for all virtual networks in an NV Domain. Before sending unicast traffic to a remote TS (or TSes for broadcast or multicast traffic), an NVE must know where the remote TS(es) currently reside. When a TS attaches to a virtual network, the NVE obtains information about that VN from the NVA. The NVA can provide that information to the NVE at the time the TS attaches to the VN, either because the NVE requests the information when the attach operation occurs, or because the VM orchestration system has initiated the attach operation and provides associated mapping information to the NVE at the same time. There are scenarios where an NVE may wish to query the NVA about individual mappings within an VN. For example, when sending traffic to a remote TS on a remote NVE, that TS may become unavailable (e.g,. because it has migrated elsewhere or has been shutdown, in which case the remote NVE may return an error indication). In such situations, the NVE may need to query the NVA to obtain updated mapping information for a specific TS, or verify that the information is still correct despite the error condition. Note that such a query could also be used by the NVA as an indication that there may be an inconsistency in the network and that it should take steps to verify that the information it has about the current state and location of a specific TS is still correct. For very large virtual networks, the amount of state an NVE needs to maintain for a given virtual network could be significant. Moreover, an NVE may only be communicating with a small subset of the TSs on such a virtual network. In such cases, the NVE may find it desirable to maintain state only for those destinations it is actively communicating with. In such scenarios, an NVE may not want to maintain full mapping information about all destinations on a VN. Should it then need to communicate with a destination for which it does not have mapping information, however, it will need to be able to query the NVA on demand for the missing information on a per-destination basis. The NVO3 architecture will need to support a range of operations between the NVE and NVA. Requirements for those operations can be found in .

An NVA provides service to the set of NVEs in its NV Domain. Each NVA manages network virtualization information for the virtual networks within its NV Domain. An NV domain is administered by a single entity. In some cases, it will be necessary to expand the scope of a specific VN or even an entire NV domain beyond a single NVA. For example, multiple data centers managed by the same administrator may wish to operate all of its data centers as a single NV region. Such cases are handled by having different NVAs peer with each other to exchange mapping information about specific VNs. NVAs operate in a federated manner with a set of NVAs operating as a loosely-coupled federation of individual NVAs. If a virtual network spans multiple NVAs (e.g., located at different data centers), and an NVE needs to deliver tenant traffic to an NVE that is part of a different NV Domain, it still interacts only with its NVA, even when obtaining mappings for NVEs associated with a different NV Domain. shows a scenario where two separate NV Domains (1 and 2) share information about Virtual Network "1217". VM1 and VM2 both connect to the same Virtual Network 1217, even though the two VMs are in separate NV Domains. There are two cases to consider. In the first case, NV Domain B (NVB) does not allow NVE-A to tunnel traffic directly to NVE-B. There could be a number of reasons for this. For example, NV Domains 1 and 2 may not share a common address space (i.e., require traversal through a NAT device), or for policy reasons, a domain might require that all traffic between separate NV Domains be funneled through a particular device (e.g., a firewall). In such cases, NVA-2 will advertise to NVA-1 that VM1 on Virtual Network 1217 is available, and direct that traffic between the two nodes go through IP-G. IP-G would then decapsulate received traffic from one NV Domain, translate it appropriately for the other domain and re-encapsulate the packet for delivery.

xx xx | xxxxxxxx | xx xx +---+-+ | xx xx |NVA-1| +--+--+ xx xxx +-----+ |NVA-2| xxxx xxxx +-----+ xxxxx ]]> NVAs at one site share information and interact with NVAs at other sites, but only in a controlled manner. It is expected that policy and access control will be applied at the boundaries between different sites (and NVAs) so as to minimize dependencies on external NVAs that could negatively impact the operation within a site. It is an architectural principle that operations involving NVAs at one site not be immediately impacted by failures or errors at another site. (Of course, communication between NVEs in different NV domains may be impacted by such failures or errors.) It is a strong requirement that an NVA continue to operate properly for local NVEs even if external communication is interrupted (e.g., should communication between a local and remote NVA fail). At a high level, a federation of interconnected NVAs has some analogies to BGP and Autonomous Systems. Like an Autonomous System, NVAs at one site are managed by a single administrative entity and do not interact with external NVAs except as allowed by policy. Likewise, the interface between NVAs at different sites is well defined, so that the internal details of operations at one site are largely hidden to other sites. Finally, an NVA only peers with other NVAs that it has a trusted relationship with, i.e., where a VN is intended to span multiple NVAs. Reasons for using a federated model include: Provide isolation among NVAs operating at different sites at different geographic locations. Control the quantity and rate of information updates that flow (and must be processed) between different NVAs in different data centers. Control the set of external NVAs (and external sites) a site peers with. A site will only peer with other sites that are cooperating in providing an overlay service. Allow policy to be applied between sites. A site will want to carefully control what information it exports (and to whom) as well as what information it is willing to import (and from whom). Allow different protocols and architectures to be used to for intra- vs. inter-NVA communication. For example, within a single data center, a replicated transaction server using database techniques might be an attractive implementation option for an NVA, and protocols optimized for intra-NVA communication would likely be different from protocols involving inter-NVA communication between different sites. Allow for optimized protocols, rather than using a one-size-fits all approach. Within a data center, networks tend to have lower-latency, higher-speed and higher redundancy when compared with WAN links interconnecting data centers. The design constraints and tradeoffs for a protocol operating within a data center network are different from those operating over WAN links. While a single protocol could be used for both cases, there could be advantages to using different and more specialized protocols for the intra- and inter-NVA case.

To support peering between different NVAs, an inter-NVA protocol is needed. The inter-NVA protocol defines what information is exchanged between NVAs. It is assumed that the protocol will be used to share addressing information between data centers and must scale well over WAN links.

The NVO3 architecture consists of two major distinct entities: NVEs and NVAs. In order to provide isolation and independence between these two entities, the NVO3 architecture calls for well defined protocols for interfacing between them. For an individual NVA, the architecture calls for a logically centralized entity that could be implemented in a distributed or replicated fashion. While the IETF may choose to define one or more specific architectural approaches to building individual NVAs, there is little need for it to pick exactly one approach to the exclusion of others. An NVA for a single domain will likely be deployed as a single vendor product and thus there is little benefit in standardizing the internal structure of an NVA. Individual NVAs peer with each other in a federated manner. The NVO3 architecture calls for a well-defined interface between NVAs. Finally, a hypervisor-to-NVE protocol is needed to cover the split-NVE scenario described in .

When tunneling tenant traffic, NVEs add encapsulation header to the original tenant packet. The exact encapsulation to use for NVO3 does not seem to be critical. The main requirement is that the encapsulation support a Context ID of sufficient size. A number of encapsulations already exist that provide a VN Context of sufficient size for NVO3. For example, VXLAN has a 24-bit VXLAN Network Identifier (VNI). NVGRE has a 24-bit Tenant Network ID (TNI). MPLS-over-GRE provides a 20-bit label field. While there is widespread recognition that a 12-bit VN Context would be too small (only 4096 distinct values), it is generally agreed that 20 bits (1 million distinct values) and 24 bits (16.8 million distinct values) are sufficient for a wide variety of deployment scenarios.

The simplicity of operating and debugging overlay networks will be critical for successful deployment. Overlay networks are based on tunnels between NVEs, so the OAM (Operations, Administration and Maintenance) framework for overlay networks can draw from prior IETF OAM work for tunnel-based networks, specifically L2VPN OAM . RFC 6136 focuses on Fault Management and Performance Management as fundamental to L2VPN service delivery, leaving the Configuration, Management, Accounting Management and Security Management components of the OSI "FCAPS" taxonomy for further study. This section does likewise for NVO3 OAM, but those three areas continue to be important parts of complete OAM functionality for NVO3. As is the case for L2VPN, there is a client/server relationship between the overlay and underlay networks that is a consideration for fault and performance management - a fault in the underlay may manifest as fault and/or performance issues in the overlay. Diagnosing and fixing such issues are complicated by NVO3 abstracting the underlay network away from the overlay network (e.g., intermediate nodes on the underlay network path between NVEs are hidden from overlay VNs). NVO3-specific OAM techniques, protocol constructs and tools are needed to provide visibility beyond this abstraction to diagnose and correct problems that appear in the overlay. Two examples are underlay-aware traceroute , and ping protocol constructs for overlay networks . NVO3-specific tools and techniques are best viewed as complements to (i.e., not as replacements for) single-network tools that apply to the overlay and/or underlay networks. Coordination among the individual network tools (for the overlay and underlay networks) and NVO3-aware dual-network tools is required to achieve effective monitoring and fault diagnosis. For example, the defect detection intervals and performance measurement intervals ought to be coordinated among all tools involved in order to provide consistency and comparability of results. For further discussion of NVO3 OAM requirements, see .

This document presents the overall architecture for Network Virtualization Overlays (NVO3). The architecture calls for three main areas of protocol work: A hypervisor-to-NVE protocol to support Split NVEs as discussed in . An NVE to NVA protocol for disseminating VN information (e.g., inner to outer address mappings). An NVA-to-NVA protocol for exchange of information about specific virtual networks between federated NVAs. It should be noted that existing protocols or extensions of existing protocols are applicable.

Helpful comments and improvements to this document have come from Alia Atlas, Abdussalam Baryun, Linda Dunbar, Lizhong Jin, Anton Ivanov, Greg Mirsky, Carlos Pignataro, Dennis (Xiaohong) Qin, Erik Smith, Takeshi Takahashi, Ziye Yang and Lucy Yong.

This memo includes no request to IANA.

The data plane and control plane described in this architecture will need to address potential security threats. For the data plane, tunneled application traffic may need protection against being misdelivered, modified, or having its content exposed to an inappropriate third party. In all cases, encryption between authenticated tunnel endpoints (e.g., via use of IPsec) and enforcing policies that control which endpoints and VNs are permitted to exchange traffic can be used to mitigate risks. For the control plane, between NVAs, the NVA and NVE as well as between different components of the split-NVE approach, a combination of authentication and encryption can be used. All entities will need to properly authenticate with each other and enable encryption for their interactions as appropriate to protect sensitive information. Leakage of sensitive information about users or other entities associated with VMs whose traffic is virtualized can also be covered by using encryption for the control plane protocols and enforcing policies that control which NVO3 components are permitted to exchange control plane traffic. See the NVO3 framework security considerations in RFC 7365 for further discussion.